Temporal Networks

Petter Holme1, 2, 3 and Jari Saramäki4

1
IceLab, Department of Physics, Umeå University, 901 87 Umeå, Sweden
2
Department of Energy Science, Sungkyunkwan University, Suwon 440–746, Korea
3
Department of Sociology, Stockholm University, 106 91 Stockholm, Sweden
4
Department of Biomedical Engineering and Computational Science,
School of Science, Aalto University, 00076 Aalto, Espoo, Finland
A great variety of systems in nature, society and technology—from the web of sexual contacts to the Inter-
net, from the nervous system to power grids—can be modeled as graphs of vertices coupled by edges. The
network structure, describing how the graph is wired, helps us understand, predict and optimize the behavior
of dynamical systems. In many cases, however, the edges are not continuously active. As an example, in net-
works of communication via email, text messages, or phone calls, edges represent sequences of instantaneous
or practically instantaneous contacts. In some cases, edges are active for non-negligible periods of time: e.g.,
arXiv:1108.1780v2 [nlin.AO] 16 Dec 2011

the proximity patterns of inpatients at hospitals can be represented by a graph where an edge between two in-
dividuals is on throughout the time they are at the same ward. Like network topology, the temporal structure
of edge activations can affect dynamics of systems interacting through the network, from disease contagion on
the network of patients to information diffusion over an e-mail network. In this review, we present the emergent
field of temporal networks, and discuss methods for analyzing topological and temporal structure and models
for elucidating their relation to the behavior of dynamical systems. In the light of traditional network theory,
one can see this framework as moving the information of when things happen from the dynamical system on the
network, to the network itself. Since fundamental properties, such as the transitivity of edges, do not necessarily
hold in temporal networks, many of these methods need to be quite different from those for static networks.
The study of temporal networks is very interdisciplinary in nature. Reflecting this, even the object of study has
many names—temporal graphs, evolving graphs, time-varying graphs, time-aggregated graphs, time-stamped
graphs, dynamic networks, dynamic graphs, dynamical graphs, and so on. This review covers different fields
where temporal graphs are considered, but does not attempt to unify related terminology—rather, we want to
make papers readable across disciplines.

Contents L. Measuring inter-contact times and burstiness 14

M. Entropies and other information-theoretic
I. Introduction 2 measures 14

II. Types of temporal networks 4 V. Representing temporal data as a static graph 14

A. Person-to-person communication 4 A. Reachability graphs 15
B. One-to many information dissemination 4 B. Line graphs 15
C. Physical proximity 4 C. Transmission graphs 15
D. Cell biology 4
E. Distributed computing 5 VI. Models of temporal networks 15
F. Infrastructural networks 5 A. Models for temporal social networks 15
G. Neural and brain networks 5 1. Temporal exponential random graphs 15
H. Ecological networks 6 2. Models of social group dynamics 16
I. Other systems 6 B. Contact network models 16
C. Randomized reference models 17
III. Preliminaries 6 1. Randomized edges (RE) 17
2. Randomly permuted times (RP) 17
IV. Measures of temporal-topological structure 7 3. Randomized edges with randomly permuted
A. Introduction 7 times (RE + RP) 18
B. Time-respecting paths and reachability 7 4. Random times (RT) 18
C. Time-respecting paths with limits on waiting 5. Randomized contacts (RC) 19
times 8 6. Equal-weight edge randomization (EWER) 19
D. Connectivity and components 8 7. Edge randomization (ER) 19
E. Distances, latencies, and fastest paths 9 8. Time reversal (TR) 19
F. Average latency 9 9. Summary and guidelines 19
G. Diameter, network efficiency 10
H. Minimum spanning tree 11 VII. Spreading dynamics and compartmental models
I. Centrality measures 11 on temporal graphs 19
J. Persistent patterns 12 A. Bursty event dynamics and slow spreading in
K. Motifs 12 communication networks 20
 2

B. Burstiness and other temporal and structural (a) B

8,13,14
C B
8,13,14
C B
8,13,14
C

,11

,11

,11
inhomogeneities 21

6,7

6,7

6,7
C. Utilizing temporal structure for disease control 22 A 1,3 A 1,3 A 1,3

,7

,7

,7
4

4
2,

2,

2,
VIII. Future outlook 22 D D D
t=0 t = 6.5 t=∞
Acknowledgments 24 (b) A
B
References 24 C
D
0 5 t 10
I. INTRODUCTION
FIG. 1: Illustration of the reachability issue and the intransitivity of
To get an overview of a large, integrated system, one needs temporal networks (more specifically a contact sequence). In (a),
to zoom out from the details. For many systems, from the the times of the contacts between vertices A–D are indicated on the
Internet to the metabolism, from the proteome to the web of edges. Assume that, for example, a disease starts spreading at vertex
sexual contacts, an easy way of doing this is representing the A and spreads further as soon as a contact occurs. The dashed lines
system as a graph. A graph is a mathematical object consisting and vertices show this spreading process for four different times. The
spreading will not continue further than what is indicated in the t = ∞
of a set of vertices, the units of the system, and a set of edges,
picture, i.e. D cannot get infected. However, if the spreading started
the pairs of vertices that are interacting with each other. Usu- at vertex D, the entire set of vertices would eventually be infected.
ally, such networks are the infrastructure of some dynamical Aggregating the edges into one static graph cannot capture this effect
system—data-packet traffic on the Internet, disease spreading that arises from the time ordering of contacts. Panel (b) visualizes the
on social networks, etc.—and this dynamical system is what same situation by showing the temporal dimension explicitly. The
we are really interested in. The advantage of modeling the colors of the lines in (b) matches the vertex colors in (a).
system as a graph is that we can say much about the behavior
of the dynamical system without studying the actual dynam-
ics at all. We can estimate how much one part of the network the consequences of the time ordering by e.g. projecting out
influences another; how well the network is optimized with the interaction times.
respect to the dynamical system; which vertices play similar When one studies a network, it is usually not the net-
roles in the system’s operation; and so on [9, 38, 65, 112]. work itself (the vertices and edges) that is the object of study.
Sometimes such a crude modeling framework can be made Rather one wants to investigate a dynamical system on the net-
more powerful if one extends it to include additional levels work. In traditional network modeling one separates the un-
of detail, for example edge weights in weighted networks [8], derlying static network and the dynamical system on the net-
or the position of vertices in spatial networks [10]. In this work. Compared to this picture, temporal network approaches
review, we consider an additional dimension—time—and dis- moves information about when things happen from the dy-
cuss temporal networks, where the times when edges are ac- namical system to the network, the underlying structure on
tive are an explicit element of the representation. Until re- which the dynamics happen. Systems suitable to be modeled
cently, in most network studies, the time dimension has been as temporal networks are everywhere. The flow of informa-
projected out by aggregating the contacts between vertices to tion via e-mail messages, mobile telephone calls, and social
(sometimes weighted) edges, even in cases when detailed in- media is one such system that has recently attracted much
formation on the temporal sequences of contacts or interac- attention. Likewise, detailed understanding of the spreading
tions would have been available. Sometimes the solution has dynamics of some electronic and biological viruses calls for
been to segment the data into adjacent time windows where taking the properties of the underlying contact sequences into
contacts are aggregated into edges, and then study the time account. Studies of many networks in the life sciences—from
evolution of the network structure in these windows. Such activation sequences of genetic regulation to time-domain fea-
an approach does not cover all aspects of the temporal struc- tures of functional brain networks—may benefit from the tem-
ture of contact patterns. For example, the edges between ver- poral graph approach. Food webs and other networks of
tices of temporal networks need not be transitive. In static species evolve in time with environmental conditions that are
networks, whether directed or not, if A is directly connected to some extent a result of which species are present. This type
to B and B is directly connected to C, then A is indirectly con- of feedback fits the temporal-network framework. Another
nected to C via a path over B. However, in temporal networks, example is self-assembled networks of wireless devices and
if the edge (A,B) is active only at a later point in time than other distributed computing systems.
the edge (B,C), then A and C are disconnected, as nothing can In general, when is temporal networks a suitable framework
propagate from A via B to C (Fig. 1). Thus, the time order- for analysis and modeling? Just like for static complex net-
ing can matter a lot, and as we shall see below, the timings of works, the system under study should consist of agents that in-
connections and their correlations do have effects that go be- teract pairwise, so that the interactions have both some degree
yond what can be captured by static networks. Accordingly, of randomness and some regularity (i.e., there is some struc-
the main focus of this review is on methods that do not ignore ture). We also need to require similar properties for the tem-
 3

A A
B B

C D
D C
0 5 10
(a) Time (days)

Escort

Sex-buyer

(b) 0 500 1000 1500 2000

Time (days)
something

(c) 1500
Time (days)
2000

FIG. 2: The limits of applicability of aggregated contact sequences, in the context of spreading dynamics. Panel (a) shows a schematic
contact sequence (similarly to Fig. 1(b)) that would be fairly well modeled as an aggregated weighted graph assuming a standard random
contact process, as seen to the right. Panel (b) displays a real-world contact sequence involving two vertices in a real network of escorts and
sex-sellers [130]. The vertical lines show the times when the individuals are active in the data, while a line connecting the two individuals
indicates a contact between them. Both the behavior of the individuals and the activity of the edge between them are bursty, with periods of
intense activity followed by silent periods. (c) shows a hypothetical dynamics where one of the individuals in (b) gets a dose of, for example,
a pathogen from the other individual at every contact, and the concentration of the pathogen decays exponentially. If the individual becomes
sick when the pathogen concentration reaches a threshold (the horizontal, dashed line), then bursty dynamics would bring the level over this
threshold. On the contrary, for more regular contact dynamics such as those in panel (a), it would have time to decay below the threshold.

poral structures—they should not be too random or too regular low the dangerous level between the contacts. Thus for such a
in order to fit the framework. On one hand, one will always dynamical system, bursty edge activity would play a far more
lose information when projecting a temporal network struc- important role than for a system where the dynamics can be
ture to a static graph (see Fig. 1 for an illustration). On the modeled as a branching process [75], as is the case for many
other hand, in some cases, this loss of information is probably network-based models of disease spreading.
too insignificant to make up for the more complicated anal- A special case of the requirement that a system should have
ysis and modeling needed for the temporal graph approach. temporal structure for it to suit a temporal-network frame-
See Fig. 2(a) for an illustration of a contact sequence that is work, relates to time scales [43]. If the dynamical system on
fairly well modeled by a weighted graph with the assump- the network is too rapid compared to the dynamics of the con-
tion that contact times are random, with a frequency propor- tacts, or when edges are active, then there is no need to model
tional to the edge weight. The dynamical system of interest the system as a temporal network. One example is the Internet
on the network matters too—different systems can respond where the data packets travel much faster than the topology
differently to a specific temporal structure. For a thought ex- changes. In summary, if the system is temporally and topo-
periment, consider the empirical bursty contact pattern plot- logically connected in a way that affects the dynamics of in-
ted in Fig. 2(b). Assume that a contact triggers an increase of terest, then temporal networks may be an optimal theoretical
something (say, the concentration of a virus in the blood) in framework.
one of the vertices involved, which then decreases exponen- The study of temporal networks is very much an interdisci-
tially. Further, assume that the person gets sick and infectious plinary field, where much of the development has been taking
if the virus concentration reaches a critical level. Then, bursty place in parallel, seemingly without much communication be-
edge dynamics [7, 39, 76, 84, 163] could be of crucial impor- tween the different disciplines. This is reflected in a tremen-
tance for that something to propagate through the network. In dous amount of overlapping terminology—one concept can
a situation with more uncorrelated or evenly distributed times easily have four or five different names in the literature. Our
of contact, the virus concentration would have time to fall be- ambition is to give an overview of this research area in differ-
 4

ent fields. We will not try to gather the theory into one unified C. Physical proximity
framework. Instead, we hope that this review can help readers
from one discipline to read and understand papers in others, Proximity patterns of humans—data on who is close to
aware of the confusing terminologies. whom at what time—are important both for understanding
Another review of contributions primarily from computer the spread of airborne pathogens and word-of-mouth spread-
science can be found in Santoro et al. [134] and an overview ing of information. Such temporal networks have long been
of contributions from the network engineering community can inaccessible for large-scale studies. Rather, researchers have
be found in Kuhn and Oshman [83]. performed tedious fieldwork in some confined space like a fra-
In the rest of this paper, we will first discuss various real- ternity or an office [160]. Nowadays, it is fairly cheap to glean
world systems that can be modeled as temporal graphs. Then such information by using electronic devices. Here, the pio-
we go through theoretical developments, including measure- neering study was the Reality Mining project, where students
ments of temporal network structure, ways to meaningfully of Massachusetts Institute of Technology were equipped with
represent temporal networks as static networks, and studies of cell phones whose Bluetooth devices could detect their prox-
dynamical systems on temporal networks. imity to others [37]. The SocioPatterns project has developed
a platform that allows physical proximity measurements based
on wearable badges equipped with radiofrequency identifica-
II. TYPES OF TEMPORAL NETWORKS tion devices (RFID) [24]; these devices have been utilized in
measurements of dynamic and temporal proximity networks
A. Person-to-person communication
of patients [63], school children [142], and conference atten-
dees [140]. Because the human body acts as a shield for the
proximity-sensing RF signals, such sensors only record con-
Records of electronic one-to-one communication are par- tacts when the individuals are facing each other, and thus a
ticularly suitable for the temporal network approach, espe- contact can also be considered as indicative of communica-
cially in the context of the spreading dynamics of infor- tion between the individuals [24, 64, 121, 141]. Recording
mation or electronic viruses. Such data often come either of face-to-face communication events has also been realized
in the form of lists of messages from one person to an- with infrared sensor devices [145]. Another type of large-
other at a point in time, or a dialogue between two persons scale proximity data comes from hospitals where contacts be-
within a time interval. The first type contains networks of tween two patients that have been admitted to the same ward
e-mail messages [39, 61, 120, 157], mobile phone text mes- at the same time are recorded, sometimes including the med-
sages [162, 169], and instant messages and messages in online ical staff [99, 154]. Such data is important for studying the
forums [58, 95]. Phone calls are not instantaneous but have dynamics of disease outbreaks [64, 141] (such as MRSA) in
a specific duration, and can thus be considered to be of the hospitals and also protocols for wireless ad hoc communica-
second type [21, 71, 107, 116, 120]. However, in many cases, tion [121]. Similarly, for livestock disease, it is important to
call durations can be neglected and calls are assumed instanta- know the movement of animals between farms etc. [4, 158].
neous. In this context, temporal network modeling and anal-
ysis of various temporal centrality measures (see Sect. IV I)
can be used for designing strategies for containing the spread D. Cell biology
of malware in mobile devices [146].
There are a handful of systems in cell and microbiology that
can be modeled as networks [119]. Not all of these are dy-
B. One-to many information dissemination namic enough in nature that one would benefit from modeling
them as temporal networks. One of the systems that proba-
The broadcast of information to anyone that might listen, in bly fits the framework is the interactome—the set of molec-
contrast to one-to-one communication, is another type of in- ular interactions in a cell. The vertices of the interactome
formation spreading between humans that could benefit from are proteins or lighter molecules that can attach or otherwise
a temporal network approach. Typically, people have studied connect to one another to perform biological functions. Fre-
spreading events in blogs [1, 84] or microblogs (like Twit- quently, these interactions are represented as a static graph.
ter) [66, 87]. Liben-Nowell and Kleinberg’s study of chain- However, much of the biological functionality comes from
letter e-mails concerns an intermediate form of information the fact that the connections are not active all the time. For
transfer, between one-to-one and one-to-many [96]. These this reason, Przytycka et al. [126] believe that a “shift from
studies have until now focused on aggregated statistics, with- static to dynamic network analysis is essential for further un-
out much focus on the temporal effects we discuss in this re- derstanding” of the interactome. There is already a body of
view (with Ref. [96] as a bit of an exception, in that they also literature investigating the temporal aspects of protein inter-
study response times), so further investigations are called for. action and gene-regulatory networks [52, 78, 92, 151]. This
Yasseri et al. [164] take the time dimension into account in is perhaps the most natural level for temporal network ap-
an interesting way in their analysis of the circadian patterns proaches in cell biology—proteins are the workhorses of cells
of Wikipedia editorial activity: such activity patterns can be so any kind of cyclic, or otherwise dynamic, patterns have to
used to estimate the geographical distribution of editors. be done by a shift in the interaction network. One can also
 5

represent gene expression and regulatory networks as tempo- G. Neural and brain networks
ral networks [91, 92, 127, 165]. In these, the vertices are genes
that can be on (being transcribed) or not. Edges can be any Networks of neural connections represent another class of
of a number of functional relationships—-that one gene (via biological networks that may benefit from the temporal net-
feedback from RNAs or proteins) affect the transcription of work approach. There are several levels of structural and
another, or that two genes code for proteins that interact, or temporal connectivity, from the spiking patterns of individ-
that they are close to each other on the DNA, etc. ual neurons to more coarse-grained physiological or func-
Another biological network that changes over time is the tional connections between brain areas [20, 138]. For the lat-
metabolism—the set of chemical reactions that occur in a ter, there are different experimental modalities with different
healthy organism. The vertices in metabolic networks are tradeoffs regarding spatial or temporal resolution. Electroen-
molecular species that are connected if they are involved in cephalography (EEG) and magnetoencephalography (MEG)
the same chemical reaction. At any given time and subcel- measure electrical signals and perturbations of the extracra-
lular localization, only a part of the entire biochemical reac- nial magnetic fields, respectively, and have fairly good tem-
tion system is active. This situation changes with time, and poral but poor spatial resolutions. Functional magnetic reso-
temporal networks can potentially capture its dynamics [26]. nance imaging (fMRI) detects changes in regional brain ac-
However, the change comes about via alterations in the influx tivity by measuring blood oxygenation levels, and it has a
to the cell that reflects the overall state of the organism’s body, high spatial but poor temporal resolution. Regardless of the
or is controlled by genes. Both these processes are relatively experimental technique used, the typical approach is to use
slow compared to the conversion of molecules, so probably time series associated to vertices (individual sensors for EEG
a temporal-network analysis of metabolism does not need the and MEG, three dimensional regions—voxels—or larger ag-
more elaborate methods that we mention in this review. gregated regions for fMRI), and assign an edge between two
vertices at a given point in time or within some time window,
if the signals are correlated or in phase. Such networks are
called functional networks; in this abstraction, the existence
of a link represents simultaneous activation of brain areas, in-
E. Distributed computing
dicating a functional connection between them (see, e.g. [20]).
Naturally, such functional connections reflect the properties
Much of the early theoretical developments on temporal of the underlying anatomical connectivity network between
networks come from computer science. There are many dif- brain areas mediated via neuronal fiber bundles (the structural
ferent types of distributed computing systems but they all con- network), but functional links may appear between brain ar-
sist of fairly independent computational units spread out over eas that have no (or only few) direct physical connections. In
some network [44]. Since the computation runs in parallel general, in temporal brain functional networks, the temporal
to the information spreading between the units, they typically links represent the time dynamics of simultaneous brain area
need to operate with information that is of different age . To activations – while the structural substrate network is static on
study such a system theoretically, a central problem is esti- such time scales, functional link activations vary in time.
mating and controlling the age of the information that is ac- As an example of a static approach to functional brain net-
cessible to the vertices. works, De Vico Fallani et al. [35] use correlations of EEG
time series to derive a directed network in which they ana-
lyze the motifs (over-represented subgraphs of three vertices).
Such a study would be even richer in the temporal network
framework where the motifs would represent temporal sub-
F. Infrastructural networks
networks. There are, to our knowledge, only a few papers
where the time domain is directly taken into account: Valen-
Most infrastructural networks change so slowly that there cia et al. [156] study functional brain networks reconstructed
is no point in modeling them as temporal networks. Take the from MEG data with the phase-locking criterion, and show
Internet as an example: the dynamical system in question— that the functional connectivity varies with time and frequency
the flow of data packets—operates globally at a time scale of during the processing of visual stimuli, while certain network
seconds. The fastest changes to the network topology come features such as small-world characteristics are maintained
from new business agreements between subnetworks that are (see also [149]). Dimitriadis et al. [36] investigate brain dy-
already in physical contact. These happen globally a few namics as measured with EEG during mental calculations, and
times per minute, but compared to the size of the entire In- identify hubs that facilitate communication in the underlying
ternet, this is so slow that one can probably safely assume functional networks. Bassett et al. [12] monitor the evolution
it is static [40, 125]. However, for some types of transport of a brain network while the subject is learning a simple motor
networks, e.g. the air-transport network, it can be meaningful task. In addition, it would be of great interest to measure the
to apply certain temporal network concepts such as temporal dynamics of functional networks when the applied stimulus
path durations and centrality [120]. In this case, edge activa- is also time-dependent, especially with naturalistic (close-to-
tion sequences correspond to scheduled transport connecting real-life) paradigms such as watching a movie or listening to
vertices, such as individual flights or trains. music in the fMRI scanner (see e.g. Ref. [72]).
 6

(a) 1,2,4 (b) (1,3)

9 ) ,(5,11
, 5 )
2,3 (2,
5 )
5,1 (3,4),( 17
7,17 7,18) (8,

FIG. 4: Contact sequences and interval graphs. This figure illus-

5 10 15 20 25 30 35 40
trates the two fundamental temporal network representations in our
time (days) discussion—contact sequences (a) and interval graphs (b). The times
of the contacts are states next to the edges. We also visualize the con-
FIG. 3: A temporal network of zebras. The figure is adapted from tacts timelines (grey bars). In these the contacts are marked by black
Tantipathananandh et al. [150]. The data comes from Sundaresan et bars or fields and the time lines range from t = 0 to t = 20 (with t = 0
al. [143]. Each horizontal line corresponds to one individual. The to the left). In the former, contacts occur at points in time whereas
contacts between individuals are not shown; instead the clusters as the contacts are extended in time in the latter. Timeline plots like
identified by the algorithm in Tantipathananandh et al. are illustrated those in Figs. 1(b) and 2(b) are another suitable depiction of contact
by the colored squares. sequences, for contact graphs such illustrations are not as readable.

H. Ecological networks proach. An early paper on time-evolving network considered

supply networks for the manufacturing industry [33]. There
Ecological networks capture the interactions between are likely other economic systems that would benefit from
species or other categories of organisms [123, 137]. They temporal network modeling. Networks that, like citation net-
could be trophic, i.e. showing which species prey on another, works [105] are normally thought of as strictly growing, could
or mutualistic, i.e. representing how two species engage in a show temporal effects in the growth that could benefit from
relationship that benefits both. In many aspects, ecological being studied in a temporal-network framework.
networks are dynamic [34]. As an example, they can change
with the seasons as organisms go through different phases of
their life cycles (and thus have different capabilities and needs III. PRELIMINARIES
with respect to their interaction with others) [117, 155]. As
another example, at longer time scales, ecological networks The temporal networks we consider in this review can be
change through evolution. Since most methods of this review divided into two (rough and overlapping) classes correspond-
are specialized to cases where the time scale of topological ing to the two types of representations illustrated in Fig. 4.
changes of the network is not too much slower than the dy- In the first representation (Fig. 4(a)) there is a set of N ver-
namics of the network (the flux of matter between species), tices V interacting with each other at certain times, and the
studies of evolutionary effects might not require the tempo- durations of the interactions are negligible. In this case, the
ral network approach (at least in their traditional sense, cf. system can be represented by a contact sequence—a set of
Ref. [166]). For faster changes of the interaction patterns, in C contacts, triples (i, j, t) where i, j ∈ V and t denotes time.
response to environmental changes, yearly and circadian cy- Equivalently, one can represent the system by V, a set of M
cles, etc., temporal networks could provide a useful frame- edges (pairs of vertices) E, and, for e ∈ E, a non-empty set
work. of times of contacts T e = {t1 , . . . , tn }. Typical systems suit-
In population biology one also studies proximity and mo- able to be represented as a contact sequence include com-
bility networks of animals [16, 31, 101, 143, 150]; see Fig. 3. munication data (sets of e-mails, phone calls, text messages,
These are, just like human proximity networks, prime exam- etc.), and physical proximity data where the duration of the
ples of systems where the temporal dimension can affect dy- contact is less important (e.g. sexual networks). Commonly,
namical systems like disease and information spreading, and authors group the contacts happening at the same discrete
are thus apt for temporal-network analysis. In Ref. [5], dy- timestep into one graph (or “graphlet” in the terminology of
namical patterns of cattle movement were analyzed with tem- Ref. [13]) and present the temporal network as a time se-
poral networks where vertices represent premises, and edges quence of graphs. Since this representation makes it tempting
cattle movement among premises. to think of the temporal-network structure as an evolving static
network structure (which misses many of the unique points
of temporal networks), we prefer contact sequences. Further-
I. Other systems more, in many real datasets with a high time resolution there
are only a handful of edges present at a timestep among tens
The above-mentioned systems are far from the only poten- of thousands of vertices which make the graph-sequence rep-
tial applications of temporal network modeling. Probably, resentation a little odd (but this is really just a matter of taste).
the easiest way of finding more examples is to look at the In the second class of temporal networks we discuss, inter-
complex-network literature and to ask oneself if a certain sys- val graphs, the edges are not active over a set of times but
tem has enough temporal structure for a temporal-network ap- rather over a set of intervals T e = {(t1 , t10 ), . . . (tn , tn0 )}, where
 7

the parentheses indicate the periods of activity—the unprimed IV. MEASURES OF TEMPORAL-TOPOLOGICAL
times mark the beginning of the interval and the primed quan- STRUCTURE
tities mark the end. The static graph with an edge between
i and j if and only if there is a contact between i and j is A. Introduction
called the (time) aggregated graph. Examples of systems that
are natural to model as interval graphs include proximity net- The topological structure of static networks can be char-
works (where a contact can represent that two individuals have acterized by an abundance of measures (see, e.g., [32]). In
been close to each other for some extent of time), seasonal essence, such measures are based on connections between
food webs where a time interval represents that one species is neighboring nodes (such as the degree or clustering coeffi-
the main food source of another at some time of the year, and cient), or between larger sets of nodes (such as path lengths,
infrastructural systems like the Internet. Like for static graphs, network diameter and various centrality measures). When the
it can be useful to define an index function of whether a pair additional degree of freedom of time is included in the net-
of vertices is connected at a given time. This is the adjacency work picture, many of these measures need rethinking or re-
index (Ref. [23] call it “presence function”) vising. While some measures are perhaps best applied to net-
( works aggregated over chosen time periods (e.g. the time-
1 if i and j are connected at time t
a(i, j, t) = (1) dependent degree of a node can be computed as the number
0 otherwise of links activated within some time window), other properties
are directly influenced by the order of link activations. As
Just as the largest body of literature on network theory con- an example, paths that transmit anything through the network
siders simple graphs (of undirected edges that never occur need to follow time-ordered sequences of contacts, and like
twice between the same vertices, and never connect a ver- the temporal networks themselves, such paths are not static
tex to itself), we put (unless otherwise stated) some further but change in time. In this section, we will review mea-
restrictions on the time stamps of the edges. We assume sures proposed for characterizing temporal-topological struc-
that a triple of a contact sequence never occurs twice, which ture. Many of these build on the concept of time-respecting
means that we can order the contacts uniquely (first by the paths discussed in the first subsection, such as different cen-
time stamps, then by their smallest vertex index and finally trality measures. We also address some of the few methods
by their largest vertex index). Usually, we also disregard the proposed for characterizing mesoscopic features and patterns
order of the vertices in a contact—if contacts are considered in temporal networks; in this area, there is still a clear lack
directed, we will always indicate this separately. For interval of methods. We conclude by discussing measures of the tem-
graphs, we assume that there are no empty or overlapping in- poral inhomogeneities of contact sequences and information-
tervals. More mathematically speaking, consider two intervals theoretic aspects.
(ti , ti0 ), (t j , t0j ) ∈ T e ; then the following three statements need to
be true
B. Time-respecting paths and reachability
1. ti < ti0

2. t j < t0j Paths that connect nodes represent the pathways constrain-
ing the dynamics of any process taking place on the network.
3. ti < t j if and only if ti0 < t j In a static graph, a path is simply a sequence of edges such
that one edge ends at the node where the next edge of the
These definitions can of course be extended in many ways— path begins (such as A to B to C to D in Fig. 1). In order
one can think of weighted temporal networks (where a vertex for this concept to be meaningful in temporal networks, es-
or edge is associated with a time-dependent scalar), networks pecially in relation to dynamical processes, paths must neces-
where the edges take some time to traverse [19] or the con- sarily be constrained to sequences of link activations that fol-
tacts are completed only after some duration δt and should low one another in time. Thus, in a temporal graph, paths are
thus be represented as quadruples (i, j, t, δt) [120]. Of course, usually defined as sequences of contacts with non-decreasing
vertices could also be active intermittently, but usually this is times that connect sets of vertices. Kempe et al. [73] and other
reflected in the activity of edges and we will not discuss this authors [60] call such paths “time-respecting.” As an exam-
issue further. Such extensions might require modifications of ple, in Fig. 1, there are time-respecting paths from A to D
the methods and measures we discuss in this review. Some (for example (A, B, 7), (B, C, 8), (C, D, 11)) but none from A
such modifications are straightforward, while others are open to E. In the literature, the terms “journey” [19, 41] and “non-
research questions; several concepts of temporal graphs have decreasing path” [27] have also been used for time-respecting
no immediate counterpart in static graphs. paths.
There are other representations present in the literature. The constraint of having to follow time-ordered sequences
Those that reduce the information from the original tempo- of contacts gives rise to differences between temporal paths
ral graph by mapping them to a static graph are discussed in and paths in static networks. Similarly to static directed
Sect. V. A yet rarely followed path is Harary and Gupta’s networks, it might be the case that i is reachable by time-
suggestion to model temporal graphs with logic program- respecting paths from j, but j cannot be reached from i. A
ming [54]. difference between directed and temporal networks is that the
 8

paths are not transitive. The existence of time-respecting paths 1.0

(a) (b)
from i to j and j to k does not imply that there is a path from i
0.8

reachability fraction
to k, as seen in the example above – a path from i to k via j ex-
ists only if the first contact on the j−k path takes place after the
0.6
last contact on the i − j path. This is related to a fundamental
property of time-respecting paths: they, too, are temporal, and 0.4
begin and end at certain points in time. Thus, the existence of
a time-respecting path that begins at i at time t0 and leads to j 0.2
does not guarantee that such a path between i and j exists for
t > t0 ; in addition, a future temporal path joining i and j might 0.0
follow a different route. Hence, the statement that ”there is 10⁰ 10² 10⁴ 10⁶ 10¹ 10³ 10⁵
∆ c (secs) ∆ c (secs)
a time-respecting path between i and j” is ambiguous; such
paths always take place within some time window.
FIG. 5: The reachability ratio as a function of the maximum allowed
Thus, time-respecting paths define which vertices can be delay of a time-respecting path at each vertex. Panel (a) shows data
reached from which other vertices within some observation for a mobile telephone call network, (b) comes from the connections
window t ∈ [t0 , T ]. The set of vertices that can be reached of an airline system. This figure is adapted from Ref. [120].
by time-respecting paths from vertex i is called the set of in-
fluence of i. This is important e.g. for disease spreading, as
it is the set of vertices that can eventually be infected if i is cell-phone calls and passenger flights (see Fig. 5). The reach-
the source of infection. It may be useful to define a set of in- ability ratio was observed to increase rather sharply around a
fluence at the time t as the set of vertices that can be reached characteristic time of about two days for the cell-phone data
via time-respecting paths from vertex i that begin at time t or and 30 minutes for the airline network. These time scales can
later. Holme [59] calls the average fraction of vertices in the be interpreted as reflecting some fundamental property of the
sets of influence of all vertices as the reachability ratio. system. Mobile phone call sequences are known to be bursty
Reversely, one can also define the source set of i as the and this is reflected in long inter-call intervals, and thus in-
set of vertices that can reach i through time-respecting paths formation must still be further transmitted two days after its
within the observation window. This set consists of all ver- reception if it is to reach a large number of individuals. In-
tices that can have been the source of an infection infecting terestingly, 30 minutes is about the minimum allowed transfer
i. Riolo et al. [128] points out the size of i’s source set—i’s time between flights for transferring passengers.
source count—as an important quantity. Moody [108] gives
another definition of the reachability of a vertex that increases
in proportion to the count of time respecting paths. Again, D. Connectivity and components
as the source set is time-dependent, one may also monitor the
source count a function of time, i.e. study how many other
Connectivity—whether or not a pair of vertices is con-
vertices may reach vertex i by time-respecting paths by time
nected by a path—is a fundamental concept for networks. Any
t0 , when the paths begin no earlier than t < t0 . It may be use-
network can be divided into sets of nodes based on their con-
ful to view the two time-dependent sets – the source set and
nectivity; these sets, in turn, impose limitations on any dy-
the set of influence – as the past and future ”light cones” for
namics taking place on the network. For static networks, ver-
vertex i, i.e. the set of nodes which may have influenced i’s
tices are either connected or not, and connected components
current state and the set of nodes which may be influenced by
are defined as sets of vertices between which some path can
i in the future via time-respecting paths (see also Sect. IV D
always be found. As mentioned above, connectivity is not
below).
a symmetric relation for directed or temporal graphs. In di-
rected graphs, the property of connectivity can be divided into
two parts: strong connectivity, where there is a directed path
C. Time-respecting paths with limits on waiting times between all pairs of vertices, and weak connectivity, where
there is a path between all pairs of vertices if the edges are
Pan and Saramäki [120] note that some spreading or trans- considered undirected. These two concepts can be general-
port processes that follow time-respecting paths set limitations ized for temporal networks. Nicosia et al. [113] propose the
on the times that the paths are allowed to spend at vertices, i.e. following definitions: two vertices i and j of a temporal net-
times between two consecutive contacts on a path. These lim- work are defined to be strongly connected if there is a di-
itations may be from below—such that the process must be rected, time-respecting path connecting i to j and vice versa,
allowed to wait for some time before the next contact on the while they are weakly connected if there are undirected time-
path—or from above, as for spreading dynamics, where the respecting paths from i to j and j to i, i.e. the directions of the
transmission has to happen quickly enough before infectious contacts are not taken into account. On the basis of these def-
nodes recover, or before nodes lose their interest in forward- initions, one may then define strongly or weakly connected
ing some piece of information. They set a limit for the latter— components of the temporal graph as sets of vertices where
the maximum allowed waiting time at a vertex—and measure each pair fulfills these criteria. Nicosia et al. also show that
the reachability ratio as a function of empirical networks of the problem of finding strongly connected components in tem-
 9

poral graphs can be mapped into the problem of finding max-

8
imal cliques in affine graphs, where an element of the ad-
7
jacency matrix indicates strong connectivity between the re-
6
spective vertices in the temporal graph. One should keep in
5

latency
mind that such properties always depend on the time of mea-
4
surement, i.e. are in general only valid within some specified
3
time window. In addition to strong and weak connectivity,
2
one can define yet another type of connectivity—transitive
1
connectivity—for temporal networks. A subgraph is transi-
tively connected if time respecting paths from i to j and j to k 0 5 10
t
implies a time respecting path from i to k.
FIG. 6: The forward latency for paths from A to C of Fig. 1 as a
function time. There are two paths joining A and C that go through
E. Distances, latencies, and fastest paths an arbitrary set of vertices—the first contact of the first path takes
place at t = 7, and the duration of the path is one unit of time, i.e.
For static networks, the geodesic distance between two ver- the path arrives at C in one time unit. The next path begins at t = 11,
and takes two units of time to traverse. If one would use periodic
tices is defined as the length of the shortest path joining them,
temporal boundary conditions for paths between A and C, so that the
path length being defined as the number of links forming a first observed path joining them at t = 7 repeats at t = T + 7 where
path. Shortest path lengths obviously influence how quickly T = 13 is the observation period limit, then the arrow would go up to
anything can propagate between nodes, and their average and latency 10 and then fall down linearly like for early times and thereby
distribution determines the overall ”compactness” of a net- repeat a cyclic pattern. This figure is adapted from Ref. [120].
work. Evidently, when the dimension of time is added to
the picture, it is useful to define similar quantities character-
izing how quickly vertices can reach each other through time- by time t. We call this quantity i’s view of j’s information at
respecting paths; here, to the best of our knowledge, the earli- time t. Furthermore, λi,t ( j) = t − φi,t ( j) is called j’s informa-
est work is by Cooke and Halsey in the 60’s [30]. Here, some tion latency, or just latency, with respect to i at time t, and is
difficulties arise because the time-respecting paths are them- thus a measure of how old i’s information coming from j is at
selves temporal and because the observation window is always time t. Finally, the vector [φi,t (1), . . . , φi,t (N)] is called i’s vec-
finite, and choices have to be made e.g. regarding proper ways tor clock. This framework was introduced by Lamport [90]
of averaging over quantities. In addition, the nomenclature in and further developed by Mattern [104]. Note that the above
the literature has not yet converged. definition looks backwards in time; one may also define a for-
Obviously, any time-respecting path is associated with a du- ward latency (called “temporal distance” in Ref. [120]) τi j (t)
ration, measured as the time difference between the last and that measures how long it takes to reach j from i along the
first contacts on the path; note that some authors have called fastest path, when the measurement begins at time t. A bit
it the temporal path length [120]. Analogously to the shortest reminiscent to vector clocks, Panisson et al. [121] introduce
paths that define the geodesic distance, one can find the fastest what they call intrinsic time to denote the active time of ev-
time-respecting path(s) between two nodes; the shortest time ery vertex. This, they argue, is useful to study the temporal
within which i can reach j is called their latency (also ”tem- statistics of information spreading.
poral distance” [120]). As the concepts of temporal duration
and link-wise distance have been used interchangeably in the
literature, we will in the following reserve the word ”distance” F. Average latency
for measuring numbers of links, and ”duration” and ”latency”
for measuring times (see also Sect. IV F below). A natural use for durations and latencies is in characterizing
The concept of latency was originally introduced in the the overall “velocity” of the temporal network, i.e. measuring
study of distributed computation. A central problem in this how quickly vertices can on average transmit something to
area is keeping track of the age of information that a vertex each other along the contact sequences. However, taking an
has about other vertices. A quite reasonable assumption is average over the entire window of observation to get a value
that vertices in contact update each other’s information so that for the entire graph—or even only for a pair of vertices—is
after a contact, both vertices share the most recent informa- not that straightforward. The problem is related to the finite-
tion that either of them had before the contact. This scenario ness of the observation period, and the fact that there are typ-
is similar to the SI disease spreading model that we will dis- ically different time-respecting paths between the same pair
cuss later, if the disease transmission probability upon contact of vertices that begin at different points in time. Furthermore,
is set to 100%. In terms of information spreading, this contact close to the end of the observation window, time-respecting
process defines the fastest possible trajectories of information paths become rare as they do not have enough time to be com-
between vertices. pleted, and vertices may no longer reach each other. One pos-
Consider the vertex i at time t in a temporal network over sible quantity for measuring the velocity of paths in general
which information spreads. Then let φi,t ( j) denote the latest is to enumerate all fastest time-respecting paths between ver-
time before t such that information from j can have reached i tices and then compute the average duration of such paths.
 10

However, this measure would not reflect the frequency of the 1 1

paths, and would not be affected by waiting times before the (a) (b)

reachability fraction

reachability fraction
first contacts of such paths: if i and k are joined by one sin- 0.75 0.75

real
gle path that takes one unit of time to traverse or by ten paths
of the same duration, this average would equal unity in both 0.5 0.5
PT

randomized
cases. RT
0.25 RC 0.25
For the average latency, measuring how long it on average RE
takes to reach i from j, the situation is more difficult. Latency AR
0 0
varies with time with a saw-tooth pattern (cf. Fig. 6), where 200 250 300 350 70 80 90
the jumps occur at points where a new fastest time-respecting average latency (hrs) average latency (hrs)
path begins. Close to the end of the observation window, la- 1 1
tency becomes infinite, as paths no longer have enough time (c) (d)

reachability fraction

reachability fraction
to be completed. Averaging only over the period of finite la- 0.75 0.75
tency is problematic: if vertices i and j were connected only
by a single path that takes place late in the observation pe- 0.5 0.5
riod, their average latency would be high, whereas it would
be low if that path took place earlier. To account for this, Pan 0.25 0.25
and Saramäki [120] proposed a pair-specific temporal bound-
ary condition, where for every pair of vertices, the first ob- 0 0
10 15 20 25 1000 1500 2000
served path joining them is repeated once when computing average latency (hrs) average latency (hrs)
the latency for that specific pair. Another possibility would
be to periodically repeat the entire temporal contact sequence FIG. 7: Average latency and reachability ratio of some empirical
as was done e.g. in Ref. [71]. However, this procedure may contact sequences. This figure is reprinted from Ref. [59]. Each
give rise to artifacts and connect pairs of vertices that are not panel corresponds to a dataset: (a) is from contacts of an Internet
connected at all within the observation window. community; (b)–(d) comes from e-mail exchange, for details see
For long enough periods of observation, another difficulty Holme [59]. The randomizations are Permuted Times (PT), Ran-
is posed by the dynamics of vertices entering and exiting the dom Times (RT), Random Contacts (RC), Randomized Edges (RE)
system. If only edge activation sequences are observed, such and All Random (AR).
vertex dynamics cannot generally be distinguished from edge
dynamics—e.g. in a temporal network spanned by telephone
calls, even if a person makes only infrequent calls, one can-
not generally assume that the person wasn’t a subscriber to
any pairs of vertices defines the diameter of the network, and
the operator before the first call, or has left the operator after
the average over inverse path lengths of all paths the efficiency.
the last call. However, in some cases external information on
One option of defining the diameter in a temporal graph would
vertices is available, and then one may choose to only include
be to take the longest average latency (although one could
vertices that are known to be part of the system for the entire
again argue that one should avoid the mix-up between quan-
period.
tities whose names relate to length but that are measured in
Finally, as a word of caution about the nomenclature, as al-
units of time). Again, one would then have to choose how
ready mentioned above, some authors use the terms “distance”
to deal with infinite latencies, i.e. pairs of vertices that are
and “length” as measures of time—e.g. Kossinets et al. [79]
not connected by any time-ordered path within the observa-
define the “distance” between two vertices as the shortest du-
tion window. Chaintreau et al. [25] gives another definition
ration of any time-respecting path between them. Tang et
by requiring that the diameter should be a number as small as
al. [147] calls the average time to reach (the reachable) ver-
possible such that almost surely, increasing it would not make
tices for time-respecting paths starting in a time window early
you find more pairs of vertices connected by a time-respecting
in the data “temporal path length”. To be fair, we should
path that is shorter than the diameter. The advantage with this
point out that “distance”, as in the standard static graph defi-
definition is that it does not require all vertex pairs to be con-
nition is, being a dimensionless quantity, a bit of a misnomer
nected, which is usually the case for empirical data sets.
too. Furthermore, what we call average latency has several
names: “reachability time” in Ref. [59], “temporal proximity” Tang et al. [147] propose network efficiency (the harmonic
in Kostakos [80] [171], “characteristic temporal path length” average of the latency) as a distance metric for temporal net-
in Tang et al. [147] and “temporal distance” in Ref. [120]. works. This measure combines the average latency and reach-
ability ratio of Holme [59] by a harmonic mean. It was also
mentioned in Holme [59] but discarded with the argument that
G. Diameter, network efficiency the two properties (of how many vertex pairs that are con-
nected by time-respecting paths and how fast information can
There are several quantities that characterize the compact- reach between them) carry very different information about
ness of a static network in terms of path lengths – in addition the function of the system and should rather not be mixed up
to average shortest path lengths, the largest distance between (see Fig. 7).
 11

H. Minimum spanning tree averaging it out in Eq. (3). In this case one would have to
choose t in the interval when λi,t ( j) is finite for all i and j,
The minimum spanning tree is an important concept related and integrate the sawtooth-like latency function (see Fig. 6) in
to paths in static weighted graphs. It is defined as a subgraph this interval, or to apply the boundary-condition based method
that is a tree of minimal total weight that connects all the ver- introduced in Ref. [120]. The downside of this approach is
tices of the graph. Such minimum spanning trees are impor- that there is potentially interesting information in the intervals
tant for engineering applications as it they the cheapest way with infinite latencies, and it may therefore be better to use
to reach all vertices, when the cost of an edge is measured in other centrality metrics. An augmented closeness centrality
terms of weight. Gunturi et al. [49] presents a way to identify measure based on reciprocal latencies in the same vein as the
a related concept in temporal networks that they call “time- above-mentioned “efficiency” can be defined as
sub-interval minimum spanning tree”. This is roughly speak- 1 X 1
ing the spanning tree in an interval that has the smallest aver- C E (i, t) = (4)
N − 1 j,i λi,t ( j)
age latency.
As an unusual application of reachability analyses we men- where λi,t1( j) is defined as zero if there are no time-respecting
tion Ref. [86] that uses a temporal network approach to study paths from j to i arriving at time t or earlier [120, 134]. A
how to navigate hot-air balloons over a planet with a pre- similar definition can be used for closeness centrality based
dictably changing wind field. In this case the vertices are on pairwise latencies averaged over the observation window;
cuboid volumes of the atmosphere connected if a balloon has if there are no paths at all between i and j, the average latency
the possibility of moving from one to another. (Normally, one is infinite for this pair and the reciprocal latency is defined as
probably needs more complex topologies to benefit from a zero [120].
temporal-network approach.) The betweenness centrality C B [38, 65] is another impor-
tant centrality measure based on shortest paths, measuring the
fraction of shortest paths passing through the focal vertex (or
I. Centrality measures
edge). For static networks, betweenness centrality is formally
defined as
In network theory, numerous centrality measures have been
i, j,k νi ( j, k)
P
defined for identifying important vertices beyond the degree, C B (i) = P (5)
e.g. with respect to their average distance to other vertices i, j,k ν( j, k)
or importance for shortest paths connecting other vertices. where νi ( j, k) is the number of shortest paths between j and
As when translating other quantities from static to temporal k that pass i, and ν( j, k) is the total number of shortest paths
graphs, there is usually no unique candidate for the tempo- between j and k. This definition is straightforwardly general-
ral version of a static centrality measure. A rather straight- izable [148] to temporal networks by adding a dependence on
forward approach, however, is to replace the role of paths time t and counting the fraction of shortest or fastest time-
in static networks by time-respecting paths. In the words of respecting paths that pass through the focal vertex. In the
Moody [108]: “time-dependent centrality measures that make first alternative, the focus is on paths that contain the small-
use of the number and length of time-[respecting] paths are est number of contacts, and in the second, on paths that are
obvious choices.” The closeness centrality CC [112] is for fastest to traverse. Here, one also has to define the tempo-
static networks defined as ral boundaries—options are e.g. to use the whole observation
N−1 period, or only consider the fastest time-respecting paths that
CC (i) = P , (2) begin at times closest to t at each vertex j. Just like the reg-
j,i d(i, j)
ular betweenness, the focus on shortest or fastest paths alone
where d(i, j) is the geodesic distance between i and j, i.e. the seems rather far from the situation in many real systems, es-
closeness centrality measures the inverse total distance to all pecially since the path durations are continuous and there may
other vertices and is high for vertices who are close to all oth- be paths that are only differentially slower.
ers. Similarly, for temporal networks, one may be interested Another class of centrality measures takes its starting point
in how quickly a vertex may on average reach other vertices, in the assumption that something diffuses randomly around
and define the temporal closeness centrality as [149] the network, instead of traveling from source to target along
the shortest paths, as for closeness and betweenness. A cen-
N−1 tral vertex in such a setup is a vertex that often is occupied by
CC (i, t) = P , (3)
j,i λi,t ( j)
that something [57]. For static graphs, this approach yields
matrix-based centrality measures like the eigenvector central-
where λi,t ( j) is the latency between i and j. Just like the static ity, Katz centrality and PageRank [38, 112]. Since this review
closeness centrality is undefined for disconnected graphs, focuses on methods that do not aggregate the temporal dimen-
the temporal version does not work unless there is a time- sion, any generalization of these measures worth mentioning
respecting path from j to i ending at time t the latest. This would have to use three-dimensional tensors representing the
is, in practice, quite a hard restriction for temporal networks, temporal network. Instead of working through tensor algebra,
at least if t is rather late in the observation period. In some we describe the algorithm of a generalization of the eigenvec-
situations one may want to get rid of the time dependence by tor centrality.
 12

1. Start with a centrality value 1 at each vertex. 0.6

k
2. At every contact between vertices i and j, let the C E 0.5 C
values after the contact at timestep t be 1–γ

0.4
C Et+1 (i) = ζC Et (i) + (1 − ζ)C Et ( j) (6)

k, C, 1 – γ
and 0.3

C Et+1 ( j) = ζC Et ( j) + (1 − ζ)C Et (i). (7) 0.2

When the dataset is exhausted, we have a distribution of the 0.1

centrality that we can take as a generalization of the eigen-
vector centrality (even though it is not derived from the solu- 0
tion of an eigenvalue equation). As opposed to the shortest- 0 10 20 30 40
sampling frequency, Δ (hours)
path based measures this one becomes, automatically, time-
aggregated. The parameter ζ sets the rate of centrality trans-
FIG. 8: Three static network quantities as a function of the sampling
mitted at a contact. If ζ is large it puts a bigger emphasis on time window ∆ (from Ref. [28]). k is the average degree (twice the
recent contacts. A sensible range of ζ is the interval (0, 1/2], number of edges per vertex); C is the clustering coefficient (the num-
but in practice it is system dependent and one needs to choose ber of triangles normalized to the unit interval given the number of
it by finding the value for which it ranks of the vertices as well connected triples of vertices, see Ref. [112]); γ is the adjacency cor-
as possible compared to some external data. In the upper limit, relation coefficient of Eq. 8.
the vertices share their collective centrality upon a contact.
Similarly to the eigenvector centrality applied for directed
networks (in an iterative implementation like the one above), characteristic time scales of the behavior of the sampled peo-
the temporal eigenvector centrality algorithm can get stuck at ple. This is illustrated in Fig. 8. One of the curves in this
vertices. In other words, the centrality is not updated after figure corresponds to a nifty measure of the similarity in the
the last contact, which means the vertices whose last contact connection pattern of a vertex i by the adjacency correlation
is fairly early in the contact sequence will in the end have function (also called temporal-correlation coefficient in [149])
a score that has not been updated very recently. A remedy,
similar to the PageRank or the Katz centrality, is to add a fixed
a(i, j, t) a(i, j, t + 1)
P
centrality score to every vertex at every time step. To make the j∈φ(i,t)
γi (t) = pP (8)
j∈φ(i,t) a(i, j, t + 1)
pP
contribution uniform over time, one can normalize the score j∈φ(i,t) a(i, j, t)
after every timestep (or, equivalently, increase the added value
so that it is a constant fraction of the total centrality score). A where t represents the entire time window and φ(i, t) is the set
version of this approach is discussed in Grindrod et al. [47]. of indices j such that a(i, j, t) or a(i, j, t + 1) is one. Fig. 8
displays the degree k, the clustering coefficient C and γi (t)
averaged over i and t as a function of ∆.
J. Persistent patterns

Let us now move beyond temporal network measures re- K. Motifs

lated to paths and address temporal patterns and subgraphs.
Lahiri and Berger-Wolf [88] discuss how to identify subgraphs Another approach for detecting significant patterns in net-
that are persistent in temporal networks. They define the sup- works is related to motifs. A network motif [2] is an equiv-
port set S (G0 ) of a subgraph G0 (where G0 ⊆ Gt is a subgraph alence class of subgraphs that is overrepresented in terms of
of Gt —the graph of all edges active at time t in a temporal its cardinality with respect to some null model in a network,
graph) as the set of timesteps when G0 ⊆ Gt . They go on to i.e. a larger number of such subgraphs can be found than in
define a “frequent” subgraph that we rather call a persistent a randomized reference system [172]. Usually, the configura-
subgraph as a graph that has a support larger than a certain tion model that conserves the degree sequence but otherwise
threshold value. The authors have subsequently developed a randomizes the network is taken as the reference system. The
method of detecting periodic patterns among subgraphs using over- or underrepresentation of certain subgraphs can be re-
the support set concept [89]. lated to the function of the system, especially in directed net-
In another study of persistent patterns in a temporal prox- works where the subgraphs forming motifs can be associated
imity network, Clauset and Eagle [28] investigated time slices with e.g. information processing tasks.
of an interval graph representation as a function of the dura- There are several ways to extend this concept to temporal
tion of the time window ∆. They found that the static network networks. The easiest is to look at snapshots of the network
structure of the time slices depends much on ∆—not surpris- taken at different points in time, or alternatively aggregated
ing perhaps since the average degree grows with ∆—but some edges over a period of time, and count the different subgraphs
quantities shows abrupt changes of their ∆-scaling indicating in these snapshots. Braha and Bar-Yam [18] does exactly this
 13

for an email data set, where one snapshot network represents

contacts aggregated over a day. They look at motifs of four i 10 j 20 k i 10 j 20 k
vertices and compare them to an edge rewiring null model. 25
25
The main conclusion from Braha and Bar-Yam’s study is that
l 115 l
the denser motifs are overrepresented. This approach, how-
ever, projects out the information from the order of events,

110
125 n
as the motifs are no longer temporal networks themselves. m l 115
In Ref. [5], Bajardi et al. apply an even simpler definition

110
for their dynamical motifs that are in essence time-respecting 125 n
paths constructed from events belonging to adjacent snap- m
shot time windows. Interestingly, when the time-reversal null
model is applied on data on cattle movements (see Sect. VI C),
the number of such paths is observed to be smaller than for the FIG. 9: A contact sequence (the numbers on the edges denote the
original data, indicating the presence of an ”arrow of time” in times of contacts) involving five vertices (left), and the two maximal
the system. ∆t-connected temporal subgraphs found within this sequence when
The communication motifs addressed in Zhao et al. [168] ∆t = 15. After Ref. [81].
are also defined on the basis of static subgraphs, although
temporal information is used for defining the subgraphs of
interest. Here, the authors take an approach where commu- ral subgraphs. Temporal motifs are now defined as classes of
nication events in mobile telephone call records or Facebook isomorphic valid subgraphs, where the isomorphism is taken
wall postings are first linked to form a communication graph to include the similarity of the temporal order of events, but
if they both share vertices and succeed one another within a not their exact timings. Furthermore, for every event ei there
time limit ∆t, similarly to the path waiting-time cutoff. The is a unique maximal ∆t-connected temporal subgraph that in-
difference to the above definitions is the absence of a snap- cludes ei and the largest possible set of events that are still
shot window; temporal adjacency of events is based on the pairwise ∆t-connected. Motifs based only on maximal tem-
time difference between two events sharing a vertex, instead poral subgraphs are called maximal motifs.
of aggregating the entire system. Communication motifs are For the algorithmic solution, such temporal subgraphs may
then constructed on the basis of graph equivalence of (static) nevertheless be mapped into static directed graphs, whose iso-
subgraphs found in such networks. Zhao et al. observe high morphism can then be used for computing subgraph counts
frequencies of chain, star, and “ping-pong” subgraphs that for the temporal motif equivalence classes. One can further
can hardly be found in reference networks where the event merge equivalence classes by considering (potential) informa-
times have been randomly reshuffled (the Randomly Permuted tion flow within the event sequence of a temporal subgraph—
Times null model, see Sect. VI C). Such motifs are seen to be in some cases, there are some events whose detailed order
stable over time. Zhao et al. also define “maximum flow” mo- does not matter. When applying these definitions and a cor-
tifs, where call durations are used as a filtering criterion. responding algorithm to large-scale mobile call data, Kova-
Chechnik et al. [26] define temporal motifs for gene- nen et al. found out that temporal motifs where the event se-
regulatory networks controlling metabolic pathways. In this quences may have a causal explanation are more frequent than
case, the primary temporal unit is the genes, or vertices, not sequences where the event order appears random.
the edges. Their method proceeds by coding the time evo- Regarding temporal motifs, there are issues that still remain
lution of the expression level of a gene into a sequence of unresolved and would warrant further consideration. A very
states and comparing the relative timing of the transitions be- important issue is that of reference or null models. For static
tween theses states for different connected genes. Further- networks, motif analysis is concerned with over- or under-
more Chechnik et al. use randomization techniques similar representation of subgraphs when compared with a reference
to those in Sect. VI C to infer statistically significant “activity ensemble where structural correlations have been removed.
motifs” as they call them. However, what the reference ensemble should be for tempo-
In Ref. [81], Kovanen et al. define temporal motifs such ral motifs is far from obvious, as it is evident that for tempo-
that the equivalence classes are based on temporal subgraphs, ral graphs where the events are fairly sparse, a much smaller
i.e. the order of contacts constituting the subgraph/motif mat- number of temporal motifs can be found in a reference en-
ters. The requirements for the temporal subgraphs are based semble where the event times have been randomly reshuf-
on the following definitions: two contact events ei and e j are fled. Thus any larger temporal subgraphs where the waiting
adjacent if they share a vertex, and they are ∆t-adjacent if times between events are short are almost always overrepre-
additionally their time difference is no larger than ∆t. Fur- sented, and not much information is gained from applying the
ther, two events ei and ek are ∆t-connected, if there exists a null model. Another issue that would deserve further atten-
sequence of ∆t-adjacent events joining ei and ek . Note that tion is the recurrence of temporal subgraphs—either exactly
this sequence does not have to be a time-respecting path. A the same or to some extent similar sequence of contact events
temporal subgraph may now be defined as a set of events may take place between the same set of vertices at multiple
where all pairs are ∆t-connected. Requiring that no events points in time. Understanding such recurrent mesoscale pat-
are skipped at nodes within the subgraph yields valid tempo- terns might yield a lot of insight e.g. on the function of social
 14

communication networks, or functional brain networks. Fur- (a) B

8,13,14
C (b) B C

,11
thermore, for some systems, studying the time dependence of

6,7
the occurrence counts of temporal subgraphs might also yield
A 1,3

7
useful information.

4,
2,
D A D
L. Measuring inter-contact times and burstiness

In addition to measures characterizing the temporal net- FIG. 10: Reachability graphs. Panel (a) shows a contact sequence
works and patterns in them, it is often also useful to have (same as in Fig. 1) and (b) shows its reachability graph.
a closer look at the smallest building blocks of such net-
works – the vertices and edges, and the associated sequences
of contacts. In a typical temporal network, there are multi- (e.g. Ref. [133]) and could well be extended to temporal net-
ple contacts between connected vertices, and correlations in works. One such attempt is Timo, Blackmore and Hanlen’s
the timings of such contacts play an important role e.g. re- entropy-based method to account for temporal uncertainties
garding durations of time-respecting paths and latencies be- in communication networks [152]. While this is not a deter-
tween vertices. Especially, for temporal networks of hu- ministic temporal-network approach, as in most of this review,
man communication, it has been discovered that such tim- it points to an interesting direction for future research. Simi-
ings are often bursty and deviate from the more uniform lar ideas have been developed in the analysis of ecosystems.
times expected from a memoryless, random Poisson pro- Ulanowicz define a measure “ascendency” to quantify how
cess [7, 21, 39, 61, 69, 71, 102, 107, 115, 157, 162]. Overall, well a system process its environmental input [117, 155]. The
the burstiness manifests itself in broader-than-expected distri- measure, derived from information theory, can be adapted to
butions of inter-contact times, P(τ), defined either for nodes or ecological interactions changing in time.
their individual edges. However, because of inhomogeneities
and broad distributions of node degree, numbers of contacts
per node, and numbers of contacts per edge, directly interpret-
ing the shape of P(τ) can be difficult as it reflects combined V. REPRESENTING TEMPORAL DATA AS A STATIC
effects of burstiness and such inhomogeneities. Because of GRAPH
this, displaying scaled inter-contact time distributions has be-
come the norm [21, 71, 107]. Here, nodes (edges) are binned
The literature on static graphs is many times larger than that
according to their numbers of contacts, and the inter-contact
on temporal graphs, for a natural reason: it is usually much
time distributions are scaled by the average inter-contact time
easier to analyze static graphs, especially analytically. One
in each bin, P(τ/τ∗ ). When compared against similar distri-
approach to analyzing temporal graphs is thus to derive static
butions for uniformly random contact times, broad tails are
graphs that capture both temporal and topological properties
typically observed.
of the system. The most straightforward way is to accumu-
Another alternative is to directly derive a measure of bursti-
late the contacts over some time to form edges. This can be
ness for any sequence of inter-contact times. Here, the sta-
an informative approach to studying the time evolution of the
tistical properties of inter-contact times arising from a Pois-
static structure of an evolving network [22]. Typically, au-
son process form a natural point for comparison. Goh and
thors have either investigated a network-topological quantity
Barabási [45] use as their starting point the coefficient of vari-
as a function of time when every contact between a pair of
tion, defined as the ration of the standard deviation of the inter-
vertices that has not been in contact before adds an edge to
contact times to their mean, στ /mτ . For a Poissonian contact
the graph (e.g. Ref. [58]), or they have divided the time into
sequence, στ /mτ = 1. Using this quantity, the burstiness of a
segments and studied the structure of contacts accumulated
sequence is then defined as
over those segments [133]. As mentioned above, this trivial
(στ /mτ − 1) (στ − mτ ) way of projecting out the temporal dimension can discard in-
B= = . (9) formation and this review focuses on methods that attempt to
(στ /mτ + 1) (στ + mτ )
capture it. Nevertheless, it is useful whenever the topological
For finite contact sequences, the variance is always finite, and aspects are more important than the temporal. It may also be
B ∈ (−1, 1), such that B = 1 indicates a most bursty sequence, possible to combine these aspects: Miritello et al. [107] have
B = 0 a sequence with Poissonian inter-contact times, and proposed a way of mapping the dynamic SIR model to a static
B = −1 a completely periodic sequence. edge percolation model, where the edge weights of a network
whose structure equals that of the aggregated graph are de-
fined in a way that takes into account the temporal correlations
M. Entropies and other information-theoretic measures and inhomogeneities of edges (see Section VII). There are
other ways of encoding the temporal network structure into
Information theory is the branch of science exploring the a static graph different than the aggregated graph that incor-
limits of storing, communicating and compressing data. It has porate more temporal information than a plain projection of a
already found applications in the network-clustering literature contact sequence or interval graph into the graph dimension.
 15

A. Reachability graphs disease. The advantage over a pure line graph is that transmis-
sion graphs encode the directionality arising from the order of
An alternative graph representation that might be useful at non-concurrent relationships. However, in common to line
least for very sparsely connected contact structures is reacha- graphs of concurrent relationships, the transmission graphs
bility graphs, or “path graphs” [108], or “associated influence cannot handle edges where one vertex manages to not catch
digraph” [27]. In such a case one puts a directed edge from the disease. Furthermore, paths in transmission graphs do not
vertices A to B if there is a time-respecting path from A to B have to be time respecting. Riolo et al. [128] define some dif-
(see Fig. 10). Such a graph, thus, shows which vertices can ferent versions of their transmission graphs cutting the contact
possibly affect which others. The average degree k of a reach- sequence in various ways, but essentially the idea is outlined
ability graph is thus the average worst-case outbreak size mi- above.
nus one. In other words, for any contact structure that supports
a pandemic, the reachability graph will be dense (k ∼ N). This
is a drawback for reachability graphs since most methods to VI. MODELS OF TEMPORAL NETWORKS
analyze networks are developed for sparse graphs. Neverthe-
less, one can imagine very sparse connection structures that In general, models of networks can serve many purposes.
can benefit from such a representation. They may explain the emergence of salient network charac-
One work using reachability graphs is Bearman et al. [14] teristics, or serve to produce synthetic networks with desired,
that studies dating between high-school students. Their data tunable characteristics that can then be used in computational
fits the interval-graph framework (Fig. 4b) where contacts can experiments of dynamic processes. Another class of models is
happen at any time during the intervals, and the only thing that of randomized reference networks, where empirical net-
one knows is that the contact happens at least once. In such a works are taken as inputs and randomization procedures are
case the reachability graph is not unique and the authors plot used to remove some of their characteristic correlations. For
a maximal and minimal reachability graph. Going in the op- temporal networks, the number of models proposed in the lit-
posite direction, one can prove that given a reachability graph, erature is still fairly limited. Below, we will first discuss some
one can always construct a temporal graph that has the given models proposed for temporal social networks and epidemi-
reachability structure [27]. ological contact networks, and then move on to randomized
reference models.
Note that the word ”model” has a slightly different meaning
when used in the context of statistical inference – here, tech-
B. Line graphs
niques of inference and machine learning are used to extract
statistical models, typically from limited amounts of data.
Although it is a bit outside the scope of this review, we Methods for statistical inference of temporal networks are still
mention that line graphs have been used to study disease rare, although this problem is an important one for systems
spreading in temporal networks represented by aggregated biology, e.g. in the context of temporal networks of gene reg-
time slices [97]. A line graph of a simple static graph G is ulation, where Ref. [92] use a Markov Chain Monte Carlo
a graph whose vertices are the edges of G that are connected approach to infer the parameters of a so-called Bayesian Net-
if they share a vertex in G. Sometimes the line graph is called work representation of gene regulatory networks [91]. This
“interchange graph” or “dual graph” (the latter being a bit of approach could probably have a wider use for systems also
a misnomer since the dual of the dual of G is not G). The outside of biology.
point of studying line graphs in epidemiology is that they are
closely related to the structure of concurrent partnerships—
e.g. the number of edges of the line graph is the number of A. Models for temporal social networks
concurrent partnerships in the original graph.
1. Temporal exponential random graphs

C. Transmission graphs The method of exponential random graphs (see, e.g. [129])
is commonly used by social scientists. The parameters of such
Riolo et al. [128] present a version of line graphs adding graphs, inferred from empirical data, contain information on
some more temporal information. What they call a transmis- the importance and frequency of chosen topological elements
sion graph thus goes beyond modeling the contact structure and subgraphs, such as triangles and stars. A similar modeling
and also incorporates the disease dynamics via a parameter δ framework for temporal exponential random graphs has been
that quantifies the combined incubation time and duration of put forward in e.g. Refs. [50, 53, 77]. The basic problem in
a disease. Their starting point is a graph where an edge e is this literature is that given an observed time evolution of a set
active over an interval [tstart (e), tstop (e)]. There is a directed of states of the vertices, representing some dynamical system
edge from e to e0 in the transmission graph if e and e0 share on the graph, one should determine the parameters of a time-
a vertex, tstart (e) < tstart (e0 ) + δ and tstart (e0 ) < tstart (e). See the varying exponential graph model that essentially gives the
illustration in Fig. 11. The idea with the extra δ is to identify probability of a contact to happen between a pair of vertices
the last possible time a member of an edge can transmit the at a given time. Just like its static counterpart—exponential
 16

A B
δ
(10,20)
B C
B C B C
B F C D

(2
,10
A B

,4
C D

(1

)
(15,23)

2)
A D

,2
C F C F D E

(2

2)
D E

,1
B F

(9
(4,8) E F
F E E F
0 5 10 15 20
(a) (b) time (c)

FIG. 11: Transmission graphs. Panel (a) shows an interval graph representation of a temporal network (where each edge has only one interval,
as required by the definition by Riolo et al. [128]). Panel (b) gives slightly reduced picture of the system, where a line corresponds to an edge
in (a) and the active interval is indicated. The derived transmission graph is illustrated in (c).

random graphs—has a connection to the Ising model, the tem- tures for disease simulations. It works by selecting two edges
poral exponential random graphs can also be boiled down to with some probability every time step and swapping them as
finding a partition function, in this case a time-varying one. in step 3 of the RE procedure. As the model is based on
When this is done, the temporal exponential random graph rewiring an existing network, the topology of the accumulated
model can be used both as a reference model for measuring network and the contact patterns across edges have to be de-
biases in quantities of topological and temporal structure, as termined using some other models.
well as a generative model for tuning the structure of con- In order to generate synthetic sexual contact networks,
tact sequences for simulations of dynamical systems on top of Kretzschmar et al. [82] have proposed a model that gener-
these. ates interval graphs and allows tuning the assortativity of the
resulting networks.. The model rules are as follows:

2. Models of social group dynamics 1. (a) A new partnership is formed with probability ρ.
The individuals participating in the pair are cho-
Zhao, Stehlé, Bianconi and Barrat have presented a frame- sen according to the mixing function φ:
work for modeling social networks [139, 167] where edges i. draw two random individuals i and j;
represent ephemeral social ties, such as being in face-to-face ii. decide whether they form a pair according to
contact. Their approach is based on master (or “rate”) equa- φ(i, j);
tions that represent the expected change in the number of peo-
iii. if yes, done; else go to i.
ple in a group of a certain size, and can capture observations
like “the longer an agent interacts with a group, the less it is (b) Repeat step 1a N/2 − P times.
likely to leave the group; the more the agent is isolated the less
likely it is to interact with a group.” In essence, this framework 2. In every pair consisting of a susceptible and an infected,
deals with interval graphs rather than contact sequences, much the disease is transmitted with probability η.
like the neighborhood exchange model discussed below.
3. Every pair splits up with probability σ.
In contrary, the model by Jo et al. [68] explicitly accounts
for the contact timings, combining their short time scale with The mixing function is introduced to be able to tune the mix-
the longer time scale of network evolution. Jo et al. pro- ing by degree [112]. It could be
pose a social network model that combines features of earlier
social network models (focal and cyclic closure, tie strength ki k j
reinforcement [85]) with triad interactions and social interac- φ(i, j) = 1 − ξ + ξ 2
(10)
kmax
tion task execution from a priority queue, along the lines of
Ref. [7]. This model gives rise to networks with communities for assortative mixing (where there is a tendency for high-
of strong ties connected by weak links; the contacts mediated degree vertices to connect to other high-degree vertices and
by the links are bursty. low-degree vertices to low-degree vertices), or

(ki − k j )2
B. Contact network models φ(i, j) = 1 − ξ + ξ 2
(11)
kmax

As a simple way of extending static graphs to involve a to create disassortative mixing where high-degree vertices
turnover of neighbors is was proposed by Volz and Mey- tend to connect to low-degree vertices. The parameter ξ sets
ers [159]. This model is rather similar to the RE randomiza- the strength of the assortativity or disassortativity. kmax is an
tion procedure discussed below, but with the purpose of mim- upper limit of the degree—this was before the finding that sex-
icking the change of partnerships to generate contact struc- ual networks have power-law degree distributions (Liljeros et
 17

al. [98]). A modern approach would probably be to draw de- tions changes the dynamics a lot, then obviously those play an
grees from a distribution and then connect them with a func- important role for the dynamics.
tion like φ. Except controlling for assortativity, one can also Below, we review temporal null models introduced in the
choose φ to create serial monogamous relationships by letting literature, essentially following Holme [59] and Karsai et
φ = 1 if ki = k j = 0, otherwise φ = 0. In a paper on measur- al. [71]. For this section, we assume that the temporal net-
ing concurrency (the temporal overlap of partnerships) Morris work is a contact sequence. Some of the methods work for
and Kretschzmar [109] used the parameter values: N = 2000, interval graphs too; others can be modified to interval graphs
ρ = 0.01/day, σ = 0.001/day and η = 0.1/day. We note quite straightforwardly. In the end of the section, we sum-
that, as a model for sexual networks this model fails to cap- marize and provide some guidelines for choosing reference
ture that overall sexual activity decreases with the number of models.
partners [114].
A fair amount of subsequent works has followed ideas sim-
ilar to the stochastic pair-formation model. We will not review 1. Randomized edges (RE)
all of them but mention the “dynamic random graph models
with memory” of Turova [153], which effectively models the This method is similar to the configuration model for static
same type of time-evolving graphs as above, but use a frame- graphs mentioned above, with the additional ingredient that
work more tractable for analytic calculations. Other recent contact sequences of edges follow the edges when these are
modeling evolving interval graphs but including more struc- rewired. Algorithmically, the method is defined as follows:
ture are presented in Refs. [135, 136].
1. Go over all edges sequentially.
2. For every edge (i, j), pick another edge (i0 , j0 ).
C. Randomized reference models
3. With a probability 1/2 replace (i, j) and (i0 , j0 ) by (i, j0 )
For static networks, a common way of assessing the impor- and (i0 , j), otherwise replace them by (i, i0 ) and ( j, j0 ).
tance, unexpectedness, or over/underrepresentation of topo- 4. If the move in step 3 created a self-edge or multiple
logical features of empirical networks is to compare the fea- edge, then undo it and start over from step 2.
tures against some reference model where the network is ran-
domized. The most widely applied reference model is the con- The times of contact over an edge are kept constant. Note that
figuration model, where the links of the original network are the two alternatives in step 3 where one is randomly selected
randomly rewired pairwise. This reference model preserves are needed to remove spurious correlations if the data struc-
the original degree sequence but yields otherwise maximally ture that is used returns the vertices of an edge in a specific
random networks. Then, one can assess the significance of order; otherwise one would keep the number of times a vertex
quantitative topological characteristics of the empirical graph appears in the first argument conserved, which in practice can
by either a direct comparison to averages in the randomized give quite big differences for empirical graphs, whether the
reference ensemble, by computing Z-scores for the character- graph is small or not. To speed up the process, one can skip
istics with respect to their distributions in the reference ensem- edges that already have been rewired in step 1 (by being se-
ble given by the reference model, or by measuring the extent lected in a previous step 2). On the other hand, this procedure
to which the dynamics of some processes differ when run on is linear in M and rarely a computational bottleneck.
the empirical networks and the reference ensemble. This null model can be used to study the effect of the net-
For temporal graphs, a similar approach can be applied: in work topology, that is, the wiring diagram of the original net-
this case, the original event sequences are randomized or ran- work. The model also assumes that it is the edges rather than
domly reshuffled in a chosen fashion to remove time-domain the vertices that govern the times of contacts—after the ran-
structure and correlations. However, there are several kinds of domization procedure, both the numbers and timings of con-
possible temporal correlations and several time scales where tacts for each vertex will have changed; however their degrees
the correlations are important, and thus no single, general- in the aggregated network are retained. As the contact se-
purpose null model can be designed (the temporal config- quences follow their edges when rewiring, all temporal corre-
uration model). Rather, by designing appropriate null mod- lations and inhomogeneities associated with individual edges,
els, one may switch off selected types of correlations in order such as burstiness and the distribution of inter-contact times of
to understand their contribution to the observed time-domain edges, are retained, as is the overall event rate at every point
characteristics of the empirical temporal network. Such tem- in time. The RE procedure is illustrated in Fig. 12.
poral null models have also been applied for studying the ef-
fects of various kinds of correlations on dynamical processes
(such as spreading) on temporal graphs. A typical use for 2. Randomly permuted times (RP)
such models in studies of dynamical processes would be to
essentially apply all of them, and by monitoring how the dy- As a temporal counterpart to the configuration model, one
namics of the process depends on the reference models, to can permute the contact times randomly while keeping the
pinpoint the role of different temporal and topological corre- network structure and the numbers of contacts between all
lations on the process – if removing a certain type of correla- pairs of vertices fixed. Technically, this is much simpler than
 18

(a) A
8,13,14 B
B C

,11
C

1,
11
6,7
A 1,3 D D

7
4,
2,
E

8
F E F
4,10,11
0 5 t 10

(b) A
4,7,11 B
B C
,11

10
1,4

,14
A 8,11 D D
13
8,

E
2,

1
F E F
3,6,7
0 5 t 10

(c) A
6,7,11 B
B C
,11

C
10

8,13,14
4,

D A D
1,11
E
2,
4,
7

E 1,3 F F
0 5 10
t

FIG. 12: Illustration of two types of randomization null-models for contact sequences. (a) shows a contact sequence (the same as in Fig. 1).
In (b) it is randomized by the Randomly Permuted times procedure such that contacts happen the same number of time per edge, and the
aggregated network topology is the same. In (c) the contact sequence in (a) is randomized by the Randomized edges (RE) procedure. With
RE, the time sequence of the contacts along an edge is conserved, and so is the degree sequence of the original network, but all other structure
of the topology is destroyed. (The latter statement is perhaps not so well illustrated by this figure as there are not so many graphs with the
degree sequence of the original, aggregate graph.)

applying the edge rewiring scheme discussed above and it tions and all temporal correlations with the exception of the
only requires randomly exchanging the time stamps of all con- overall rate of contacts (such as daily/weekly pattern) have
tacts, or just randomly reshuffling the order of time stamps in been destroyed.
an array or vector. No checks similar to step 4 of the RE
rule need to be performed for contact sequences; for interval
graphs, one should check for non-overlap. As this null model 4. Random times (RT)
retains all network structure and the number of contacts for
each edge, its application can be used to study the effects of
The RP ensemble conserves the set of times of the origi-
the detailed order of events, including burstiness, inter-contact
nal contact sequence. Hence, although it destroys burstiness
time distributions of contact sequences on vertices and edges,
of events on individual vertices and edges as well as correla-
and correlations and triggering effects between contacts on ad-
tions between events such as triggered chains, the aggregated
jacent edges. The model also retains the overall rate of events
rate of events in the network is unchanged and will still fol-
in the network at every point in time, such as daily or weekly
low the typical circadian and weekly patterns of human activ-
patterns in communication networks. An illustration of RP
ity [58, 67, 94, 102, 103, 170]. So far, results indicate [71]
can be found in Fig. 12.
that such overall modulation of the event rate does not appear
to matter at least for the simple SI spreading model, whose
dynamics is dominated by edge burstiness (see Jo et al. [67]
3. Randomized edges with randomly permuted times (RE + RP) for a method that removes the contribution of such circadian
patterns). However, there are cases where such patterns might
This null model works as follows: first, the network struc- play a role, such as more complicated spreading models (e.g.
ture is randomized using the RE procedure. Then, the time SIS or SIR). The random times (RT) null model destroys such
stamps of all contacts are reshuffled with the RP scheme. Thus patterns: in this approach, each contact on each edge is as-
the outcome is a temporal graph, where all structural correla- signed a random time within the observation time window of
 19

the original data, and thus the network structure and the total contacts) in the aggregated network spanned by the contact se-
number of events on each edge are conserved. When compar- quence, and thus weight-topology correlations are destroyed
ing the results of any dynamics taking place on the RP and in the null model. However, the inter-contact time distribu-
RT models, the difference should then be due to the network- tions of contact sequences of edges are not changed—the se-
level temporal patterns of the empirical data. Note that an- quences are just placed elsewhere in the network.
other alternative for uniformly random contact times is gen-
erating them from any chosen distribution or process, such as
the Poisson process [71], with parameters set up so that the 8. Time reversal (TR)
numbers of contact per each edge are on average conserved.
The above approach assumes the degree distribution is ex- This null model is designed for assessing the frequency and
ternal, as it is only investigating ensembles where the topol- importance of causal sequences [5] of contacts, where con-
ogy and the degree of each vertex is conserved. Another op- tacts trigger further contacts, and simply involves running the
tion would be to run the dynamics on Poisson random graphs original event sequence backwards in time. If sequences of
(Newman [112]) with the same number of vertices and edges consecutive contacts would be caused by temporal correla-
as the original graph and time stamps distributed as the origi- tions alone, similar numbers of such sequences should be ob-
nal data, permuted like RP and random like RT. served when time runs forwards and backwards; the lack of
such chains in the time-reversed null model compared to the
original sequence can be attributed to “the arrow of time”.
5. Randomized contacts (RC)

Here, one keeps the graph topology fixed but redistributes 9. Summary and guidelines
the contacts randomly among the edges. After this random-
ization, the number of contacts per edge follows the binomial The different randomized reference models discussed above
distribution rather than some broad, right-skewed distribution retain and destroy specific kinds of topological and temporal
as is typical for empirical data. This randomization can be correlations, and thus e.g. in studies of dynamical processes,
utilized in testing the effect of the distribution of the number they allow for pointing out the importance of various corre-
of contacts per edge in combination with the order of events. lations: the most important correlations can be pinpointed by
Hypothetically, one would like to test the effect of the distri- comparing the effects of different randomization models on
bution of the number of contacts alone, keeping the structure the dynamics. The RE and RP models permute edges and
of the temporal order of the real data. For example, a vertex contact times. Their simultaneous application (RE+RP) de-
that is active primarily in the early stage of the data would be stroys all correlations except for patterns in the overall con-
so in the randomized data. This would need a more elaborate tact rate – this provides a good starting point for the limiting
approach. case of uncorrelated temporal networks, and by additionally
randomizing contact times (RT), the overall patterns are also
removed. When studying the roles of the exact contact tim-
6. Equal-weight edge randomization (EWER) ings on edges and the correlations between adjacent edges,
comparing the EWER and ER models to the RT model should
Karsai et al. [71] use a special null model that is de- do the trick, as the static network features are retained except
signed for removing timing correlations between the contact for weight-topology correlations removed by the ER.
sequences of adjacent edges, while retaining temporal char-
acteristics associated with edges, including the distribution of
inter-contact times on individual edges. Whole contact se- VII. SPREADING DYNAMICS AND COMPARTMENTAL
quences associated with edges, i.e. all contacts and their time MODELS ON TEMPORAL GRAPHS
stamps, are randomly exchanged between edges that have the
same number of contacts. Thus single-edge patterns, such as One of the key insights of network theory is that the un-
burstiness, are retained, together with all properties retained derlying network structure can strongly affect dynamic pro-
by the RP model (number of contacts on each edge, network- cesses that are mediated by the edges. This is especially
level event frequency patterns, topological structure). This important for spreading processes, where biological or elec-
null model requires a large enough system so that there are tronic viruses, rumors, or pieces of information are transmit-
enough edges with the same number of events. ted through edges of physical contact, social ties, or electronic
connections. For such processes, the network structure affects
the speed of spreading as well as the extent of spread through
7. Edge randomization (ER) the network through features such as short path lengths [161],
the degree distribution [11, 124], degree correlations [17], or
This null model is similar to the EWER model with the ex- community structure and correlations between tie strengths
ception that the sequences can be exchanged between edges and network topology [116, 122].
that have any numbers of contacts. This corresponds to ran- Many types of spreading processes can be modeled as com-
domly exchanging the edge weights (measured as numbers of partmental models, where each vertex is in one of the charac-
 20

teristic states of the model [173], i.e. belongs to one class or stead be described with heavy-tailed or power-law distribu-
compartment [3, 56]. In the simplest SI model (Susceptible– tions. Consequently, the dynamics of spreading processes also
Infective), individuals are initially susceptible, and become differs from expectations. This failure of the Poissonian ap-
infected at some rate µ when in contact with infective indi- proximation (that is typically assumed to hold for spreading
viduals. In static networks, this process will eventually in- models on static networks) was first addressed by Vazquez et
fect all vertices that can be reached from the source, and al. [157], who studied email activity patterns from university
thus the network structure affects only the speed of spread- and service provider logs.
ing. The SIR model (Susceptible–Infective–Removed) shows Their study was motivated by the fact that the numbers of
richer dynamics. In this model, infected individuals recover new infections by computer viruses and worms that spread
and become immune to spreading, either with some probabil- through emails should decay exponentially, if the Poisson ap-
ity β per unit time or after a fixed period of time. The outcome proximation was accurate. However, reports of new infec-
of the SIR dynamics now depends on the interplay between tions are typically still published years after the release of
the two rates (infection rate µ and recovery rate β) as well as anti-viruses. In the Poisson approximation, the probability
the network structure. The process may die out, only infect- of one vertex to interact with another within a time interval
ing local clusters of vertices, or percolate through the network dt is dt/hτi, where hτi is the average interevent time. Thus,
in an epidemic fashion, such that a substantial fraction of in- the time between consecutive contacts is exponentially dis-
dividuals is infected. Beyond these two simple models, more tributed, P(τ) ∼ exp(−t/hτi). However, for the email data
complicated versions have been formulated, such as the SIRS studied by Vazquez et al., after correcting for finite observa-
model, where immunity is not permanent but individuals may tion window, it was seen that P(τ) ∼ 1/τ, followed by an
again become susceptible. exponential cutoff.
Applying such models to static networks is equivalent to The Poissonian and power-law inter-event time distribu-
assuming that all interactions between vertices take place uni- tions result in different generation times, that is, times be-
formly in time. However, in reality this is usually not the case. tween one vertex becoming infected and transmitting the in-
The spreading of biological viruses depends on temporal pat- fection to its neighbor. On the basis of generation times,
terns of contact, and the flow of information in social networks Vazquez et al. [157] calculated an analytical approximation
is influenced by the social activity patterns and rhythms of in- for the prevalence dynamics of the SI model assuming a tree-
dividuals. Importantly, there is an increasing body of evidence like structure, and showed together with simulations using
of very large heterogeneity in the timings of such interactions, empirical email sequences that the late-stage exponential de-
from burstiness in patterns of communication via electronic cay of the number of new infections is significantly slower for
and physical mail [7, 39, 61, 69, 115, 157], mobile telephone the broad inter-event time distribution and matches well with
calls and text messages [21, 71, 107, 162], instant messag- data on real computer worms.
ing [95], and patterns in proximity dynamics measured with Later, Min, Goh and Vazquez [106] studied the SI model
RFID sensors [24, 64, 141]. Daily and circadian rhythms of with power-law inter-event time distributions using theoret-
human dynamics [58, 67, 94, 102, 103] may also play a role. ical arguments that assume a tree-like network structure, as
Furthermore, there are correlations where interaction events well as with simulated spreading with uncorrelated power-law
trigger further events, such as reception of an email triggering activity patterns and the priority-queue network model. They
a forwarding event, or incoming calls causing outgoing calls concluded that a power-law waiting time distribution leads to
in mobile call networks [61, 71, 107, 120]. An overview of a power-law decay in the number of new infections in the long
dynamic processes important for infectious disease spreading time limit, with an exponent determined through the genera-
can be found in Bansal et al. [6]. tion time distribution.
The above temporal inhomogeneities, together with the fact
More empirical evidence for slower-than-Poissonian
that spreading processes have to follow the time ordering of
spreading dynamics was provided by Iribarren and
events, have important and at times drastic effects on the dy-
Moro [61, 62], who performed a viral marketing exper-
namics of spreading on temporal graphs. In the recent years,
iment with emails, where 31,183 individuals forwarded
such effects have been studied with the help of simulations
recommendations. They concluded that the large hetero-
building on empirical contact or interaction sequences, as well
geneity found in the response times is responsible for the
as with analytical tools. Although much remains unknown,
slow dynamics of information at the collective level. In
some clear conclusions can already be drawn from the exist-
their experiment, subscribers to an online newsletter were
ing literature, such as the importance of burstiness in slowing
rewarded for recommending it via email to their friends.
down epidemic-style spreading dynamics.
The viral spread of this recommendation email was tracked
at every step, providing a detailed view on reception and
forwarding events and their temporal correlations. As typical
A. Bursty event dynamics and slow spreading in for such cascades, the average number of secondary cases per
communication networks infected individual R0 was below the tipping point R0 = 1,
in this case R0 ≈ 0.26, and thus all cascades stopped in a
Human communication dynamics is almost universally finite number of steps. The cascades were observed to be
bursty, i.e. the timings between communication events devi- treelike, with a very low clustering coefficient. It was seen
ate largely from uniform or Poissonian statistics and can in- that there is a large variability to the number of forwarded
 21

recommendation emails and it was argued that this variability level call frequency is low at night and peaks around noon and
is not directly related to the degree of individuals in the early evening was seen not to contribute significantly, when
email network. Furthermore, those individuals who became investigated with a Poissonian event-generating (RT) model.
secondary spreaders, i.e. forwarded the recommendation, In addition, the EWER scheme that destroys triggering, i.e.
typically forwarded all their recommendation emails simul- temporal correlations between adjacent edges, was seen to re-
taneously in a single spreading event, and did not remain sult in slightly slower-than-original spreading for short time
spreaders for longer. The response times (times between scales, and slightly faster-than-original for long time scales.
reception and forwarding) were seen to follow a lognormal Interestingly, in contrast to these observations, Rocha et
distribution. Traditional analytic epidemic models were seen al. [131] found that temporal order and correlations speed up
to fail in predicting the speed and dynamics of the number epidemic spreading in their data set [130] of sexual contacts in
of individuals who received and forwarded the message; Internet-mediated prostitution – the origins of this difference
according to the simple exponential growth equation, most still remain unclear. Furthermore, in contrast to both findings,
new infections should happen during the early few days, Stehlé et al. found that when simulating the SEIR spreading
whereas a significant fraction of new infections was observed dynamics on a temporal contact network of conference at-
even at the time scale of months. However, the observed tendees recorded with proximity sensors of the SocioPatterns
slow dynamics was well captured by the non-Markovian RFID platform [24], the spreading dynamics is well described
Bellman-Harris branching model [55] where the lognormal by a static aggregated network if the heterogeneity of the con-
distribution of response times was used. tact durations is taken into account as edge weights.
Fernandez-Gracia et al. [46] studied the effect of broad Miritello et al. [107] also used mobile telephone call
inter-event time distributions on the ordering dynamics of the records (9 billion time-stamped calls of 20 million users over
Voter model; similarly to the slowing down of compartmen- 11 months) in their studies of simulated information spread-
tal spreading models, such distributions were observed to give ing. They applied the SIR model with a homogeneous and de-
rise to slow ordering dynamics of the model. terministic recovery time and variable transmission probabil-
ity λ with the empirical call sequence. The scaled distribution
of relay times τi j , that is, times between user i participating
B. Burstiness and other temporal and structural in a call with any other user and user i participating in a call
inhomogeneities with user j, was seen to be heavy-tailed while also display-
ing a larger number of short relay times than expected from
Karsai et al. [71] provided further insight into the effect the Poissonian case. Similarly to Ref. [71], the abundance
of temporal heterogeneities on spreading dynamics by study- of short relay times was interpreted as a signature of group
ing the behavior of the SI model. The model was simulated conversations, where calls trigger further calls and the activity
with real mobile telephone call and email contact event se- patterns of adjacent edges are thus correlated.
quences, together with reference models that destroy selected For SIR dynamics, such group conversations and corre-
correlations (see Section VI C). The largest data source was lated contact sequences were observed to give rise to larger
the call database of a mobile telephone operator, containing spreading cascades for small values of λ below the percola-
about 325 million time-stamped call records over a period of tion point than for the time-shuffled reference model where
120 days. For assessing the importance of different tempo- the distribution of relay times approaches the Poisson distri-
ral and structural inhomogeneities on the spreading dynam- bution. However, for large values of λ, the opposite behavior
ics, several reference models were used (with reference to the was observed. Thus, in the context of information propaga-
Section VI C): RP, EWER, ER, a combination of RE and RP, tion, correlations from group conversations make information
and RT with independent Poisson processes on each edge. spreading more efficient at the local small scales when a more
As in the above papers, inter-event times were seen to have realistic low transmissibility λ is used. Miritello et al. [107]
a broad distribution; these were studied by calculating the dis- also proposed a way of mapping the dynamic SIR model to a
tributions for edges binned by event numbers and rescaling static edge percolation model, similar to Newman [111] or
those by the average inter-event time in each bin, similarly to Kenah and Robins [74]. They validated their approach by
Candia et al. [21] and Miritello et al. [107]. It was also seen successfully predicting the percolation threshold for the SIR
that the scaling breaks down for small times, around 20 sec- model on empirical data. This was done approximate the
onds, indicating correlations and triggering of events. When average number of secondary infections by an effective, dy-
compared to the original event sequence, the times to full namic transmissibility—what they call the “dynamic strength
prevalence (100% of vertices infected) were seen to be shorter of ties”—obtained from the mentioned mapping.
for all null models. The RP model that destroys burstiness It is worth noting that the speed of SI spreading—especially
and correlations, except for daily patterns, gave faster spread- in the deterministic case where an infectious individual al-
ing than the ER model that destroys weight-topology corre- ways infects a susceptible individual upon contact—is related
lations but retains burstiness. Hence, the slowness of spread- to temporal path lengths, latency and reachability [59, 120],
ing can largely be attributed to the bursty event sequences on addressed elsewhere in this review. By placing constraints
individual edges, in addition to the “weak-edge” bottleneck, upon the timings between consecutive events defining a path,
i.e. edges between communities having a smaller contact fre- fastest temporal paths can also be representative of the path-
quency. However, the overall daily pattern where the system- ways taken by the SIR spreading process [120].
 22

Finally, in addition to temporal inhomogeneities slowing person gets the disease is proportional to k, and the expected
down spreading in human communication networks, a simi- number of other people the person infects is also proportional
lar effect has been observed for ants, however, compared to to k [3]. In a temporal network, it is important to remem-
a different null model. In Ref. [16], temporal networks were ber that future contacts cannot be used to determine the right
constructed based on 30-minute video recordings of ants in targets for vaccination, and that past contacts cannot bene-
6 colonies and tracking all contacts between individual ants. fit from current vaccination, unless the contacts are repeated.
Compared to a kinetic null model, where ants were consid- The basic assumption of classic network epidemiology—that
ered as gas particles that randomly collide and change di- the network is static—means that the past contacts will always
rections, SI-like information flow was observed to be signifi- also persist into the future, and this does not hold in general
cantly slower for the empirical contact sequences at long time in temporal networks.
scales. Lee et al. [93] proposed an extension of the neighborhood
We also mention the modeling work by Kamp [70], where vaccination scheme to temporal networks. They found that
the author proposes a framework to study disease spreading the strategy to ask about your most recent contact, or most
in temporal networks, which simulates the contagion pro- frequent contact some time back in the past, improves the
cess without explicitly generating interval graphs. In Kamp’s neighborhood vaccination in some real contact data sets (e.g.
setup, vertices come in to the system with a degree sampled the prostitution sex network of Rocha et al. [130], the email
from a degree distribution (which then changes due to the evo- data of Eckmann et al. [39], and the hospital proximity data
lution of the graph); they live for a limited time and their de- of Liljeros et al. [99]). The response is different for differ-
gree changes due to both the birth and death of neighbors, ent data sets, so for the hospital and prostitution data, the
and rewiring of edges. This framework allows for some ap- “most recent”-version is the most efficient, whereas for the
proximate analytical treatments with generating functions and email data the most frequent is more effective. This, Lee et
differential equation modeling but seems to require numerical al. argue, comes from the fact that the contacts along an edge
simulations for a full characterization. in the hospital proximity and prostitution data are fairly lim-
ited in time—two people who enter into a period of frequent
contacts in either one of these data sets will rather likely be
C. Utilizing temporal structure for disease control in contact with other persons a bit later. For the email data,
the driving force behind the efficiency of the ”most frequent”-
The structure of contact patterns not only affects the spread- protocol is that the contact frequency along an edge varies,
ing of disease, but this structure can also be exploited in con- but people in this dataset typically keep the relationship going
trolling and preventing the spread. The most common pre- throughout the data. From these simulations one can see that
ventive intervention is vaccination that lowers the probability the temporal structure actually adds something that can be ex-
of people catching the disease and spreading it further. Typ- ploited to local vaccination programs. We believe other types
ically, one does not have to vaccinate the entire population of population-sampling protocols that are affected by network
to block the possibility of outbreaks. Already at some partial structure in static simulations could need to be extended to
coverage f of vaccinees in the population, disease cannot any temporal networks.
longer propagate. This effect is called herd immunity. Low-
ering the threshold of herd immunity is an important goal in
public health. In static network theory, there are methods that VIII. FUTURE OUTLOOK
utilize the network topology to identify important targets for
vaccination. Perhaps the most well-known is the neighbor-
In this review, we have illustrated how several systems can
hood vaccination protocol of Cohen et al. [29] that works by
benefit from the temporal network approach, and discussed
repeating the following steps:
ways and methods of discovering and measuring network
1. Take a random person in the population. structure that lives in the time domain. Such structure be-
comes important for network studies especially in the con-
2. Ask the person to name a friend (or rather someone that text of some dynamics taking place on the network: if there
person meets regularly in such a way that the disease in are inhomogeneities and correlations in the contact sequences
question might spread). between vertices, these inhomogeneities may have dramatic
3. Vaccinate the friend. effects on dynamics mediated through the contacts. The con-
ceptual differences between modeling dynamical processes on
These steps are repeated until the desired fraction f of the static and temporal networks deserve some attention. For ex-
population is vaccinated. The benefits of this scheme are ample, for simple contact processes such as disease spread-
twofold. First, it utilizes only local information—a person is ing models, in the static network approach one typically inte-
expected to know his or her own contacts, not any third per- grates two components in the model: when the contacts along
son’s contacts—which really is a prerequisite rather than just an edge take place, and the probability for the disease be-
an advantage. Second, it samples people in proportion to their ing transmitted during a contact between a susceptible and
degree k. The importance of a person with degree k with re- an infected vertex. The common assumption is that of con-
spect to disease spreading is, in classic network epidemiology, tacts spread uniformly in time. In the temporal network ap-
proportional to k2 —roughly speaking, the probability that the proach, the first component—the timings of the contacts—is
 23

no longer a part of the spreading model, but rather an inte- structure may be, rather than first quantifying and characteriz-
gral part of the contact structure, i.e. the network itself. In ing the structural features of real systems, and only afterwards
general, for models of dynamical processes taking place on investigating the role of particular structural features on dy-
networks, moving from the static network framework to that namics. Today, we know that temporal network structure can
of temporal networks is equivalent to removing the tempo- make a difference, but not exactly how or why.
ral component related to contact timings from the model— Understanding the driving mechanisms. A third, largely
such a component is part of almost every model, although it unexplored theme is why contacts between two vertices
may not be explicitly visible—and considering temporal con- in a temporal network happen when they happen—there
tact structure instead. But is the temporal network approach are skewed inter-contact time distributions along individual
then nothing more than shifting a component of a dynami- edges, but in general, why? There are many papers about why
cal model from the dynamical system to the underlying rep- two vertices become connected by an edge [112] and there
resentation of interaction structure? What do we gain from are papers explaining time series of when a vertex does some-
such an approach, where the representation of the system in thing [7]. However, presumably, what other vertices vertex i
question necessarily becomes more complicated? First, as is connected to could affect the times when i does something,
we have seen in this review, the temporal structure is crucial and vice versa. This question comes close to the goal of adap-
for studies of dynamical processes. Luckily, at least for pro- tive network studies [48] that model the feedback from net-
cesses involving humans, much of the data produced by to- work structure (and how it affects dynamics on the network)
day’s technologies—be it mobile phone records [71, 116] or to the success of the agents forming the network (and how they
wearable sensors [24, 37]—comes in the form of contact se- seek to change their position in it). If one could include when
quences and are thus directly suitable for temporal network contacts happen along an edge into adaptive network models
studies. Thus one can go straight from the raw data to simu- and thereby explain some observed temporal-topological cor-
lations of spreading processes (or some other dynamics), and relations, this would be a breakthrough (no matter what the
analyze the role of the temporal and topological structure by objective system is).
comparing the results to reference models [131]. Second, Inference problems. Another set of more statistics-related
one may learn a lot about the system and its driving forces challenges concerns inference problems. How can one con-
by studying its temporal patterns and structure, from inter- struct a temporal network from various amounts of informa-
contact statistics to patterns involving multiple vertices such tion about the states of vertices or edges [1]? How can one
as motifs. Additionally, temporal and topological structures infer spreading chains, if one has an incomplete temporal net-
can be correlated [114], and modeling this with an underlying work? Many inference problems on static networks should
static network structure is not straightforward. be rather different on temporal graphs as many fundamental
The study of temporal networks, their characteristic fea- properties—like the transitivity of edges or Menger’s theo-
tures, and their dynamics is still a rather young field, and there rem [15]—are only valid under rather strong assumptions.
are many open questions and unexplored directions. Below, dynamical systems. If a system benefit from being modeled
we list some of these issues: as a temporal network is not only a question about the struc-
Generative models for temporal networks. There are only ture of the contacts, but also about the nature of the dynami-
very few models for temporal networks and their contact se- cal system acting over the contacts. Some dynamical systems
quences, and one of the important open issues is clearly con- might be more or less sensitive to temporal effects. In social
structing and studying parametrized, generative models of information spreading, a person may, hypothetically, spread
temporal networks, e.g. of human contact sequences with their information only if he, or she, hears about it from two inde-
characteristic features observed in real-world data, such as pendent sources within a short period of time. Such a dynami-
skewed inter-contact time distributions, bursty dynamics, and cal system should be sensitive to temporal effects like activity
circadian and weekly rhythms. bursts and the order of events.
Measures for temporal network structure. Although a large Community, cluster or mesoscopic structure. The recent
number of measures and characteristics for temporal graphs years have seen tremendous efforts for discovering meso-
have been discussed in this review, they have mostly been gen- scopic structure in static networks in the form of communi-
eralizations of static network measures, and we feel that there ties [42], loosely defined as groups of vertices more densely
is much room for improvements, e.g. in relation to simple connect within than between each other. Most of the litera-
measures of time-domain correlations of contact sequences. ture on community structure of static network focus on deriv-
As an example, why does randomizing the order of contacts ing a method for decomposing the network from some kind
in a temporal network often, but not always, make spread- of conceptually simple principle. Few, if any, studies seeks
ing dynamics slower? It would be a little breakthrough if this to identify structures known a priori to exist. The works in-
question could be answered in terms of a simple, easily ob- corporating a time dimension into the community detection
servable measure (akin to the statement “high clustering co- (like Refs. [110, 118, 132, 133]) operate on aggregated time-
efficient slows down disease spreading”) [144]. In a sense, slices of the temporal network. One can imagine clustering
recent studies on temporal networks are going in the opposite algorithms based on more elaborate temporal structures, like
direction from the early work on complex networks during the time-respecting paths (a rare exception is Ref. [100]).
millennium’s first years [112]—beginning with the effects of Visualization. Visualization software is a great help for in-
the temporal structure on dynamical systems, whatever this vestigating static networks. Even though small graphs are im-
 24

possible to embed in Euclidean space such that the Euclidean poral network framework has, with not so many exceptions,
distance between vertices is proportional to the graph distance been investigated theoretically rather than used to explain the
(that would be ideal for a direct correspondence between the world around us. Yet most complex systems in the world are
real graph—the terrain—and the visualization—the map), still time-dependent, dynamical and in motion. As we believe that
one can make visualizations that capture much of the network the temporal networks framework is really a tool for advanc-
structure. Typically they organize the vertices such that there ing science, we hope to see theoretically-minded researchers
is a positive correlation between the graph and Euclidean dis- bringing it into collaborations with their applied colleagues.
tances [51]. This is enough to see differences between net-
works of different degree–degree correlations [112] or to iden-
tify dense clusters by the eye [42]. Stacking snapshots of a
temporal network to a movie usually does a bad job to visu- Acknowledgments
alize temporal network structure, especially for sparse con-
tact sequences. The time-line plots of e.g. Figs. 1, 3 and 12 The authors acknowledge financial support by the Swedish
resolve the vertices in one dimension, which makes it even Research Council (PH), the WCU program through NRF
harder to put vertices that are close in terms of latency (or Korea funded by MEST R31–2008–10029 (PH), EU’s 7th
some other distance-like metric for temporal networks) close Framework Program’s FET-Open to ICTeCollective project
to each other in Euclidean space. If one could find another no. 238597 (JS) and the Academy of Finland, the Finnish
clever way to layout a temporal network that captured reacha- Center of Excellence program 2006–2011, project no. 129670
bility and latency that would be a very valuable contribution. (JS). The authors thank (in alphabetic order) Albert-Lászlo
These open directions mentioned above mainly concern Barabási, Ginestra Bianconi, Ciro Cattuto, Aaron Clauset,
theoretical and methodological developments. However, the Mikael Huss, Beom Jun Kim, Mario Konschake, Mirjam
real acid test of temporal networks as a fruitful paradigm is its Kretzschmar, Vito Latora, Sune Lehmann, Jure Leskovec, Ce-
application to concrete, specific problems in population biol- cilia Mascolo, Esteban Moro, Zohar Nussinov, Etsuko Non-
ogy, cell biology, ecology, neuroscience, social and political aka, John Tang, Robert Ulanowicz and Tao Zhou for com-
sciences, economics, chemistry and so forth. So far, the tem- ments.

[1] E. Adar and L. A. Adamic. Tracking information epidemics Carlson, and S. T. Grafton. Dynamic reconfiguration of human
in blogspace. In Proceedings of the 2005 IEEE/WIC/ACM In- brain networks during learning. Proc. Natl. Acad. Sci. USA,
ternational Conference on Web Intelligence, pages 207–214, 108:7641–7646, 2011.
2005. [13] P. Basu, A. Bar-Noy, R. Ramanathan, and M. P. John-
[2] U. Alon. Network motifs: Theory and experimental ap- son. Modeling and analysis of time-varying graphs. e-print
proaches. Nature Review Genetics, 8:450–461, 2007. arXiv:1012.0260.
[3] R. A. Anderson and R. A. May. Infectious diseases in human. [14] P. Bearman, J. Moody, and K. Stovel. Chains of affection: The
Oxford University Press, Oxford UK, 1991. structure of adolescent romantic and sexual networks. Ameri-
[4] P. Bajardi, A. Barrat, F. Natale, L. Savini, and V. Colizza. can Journal of Sociology, 110:44–91, 2004.
Dynamical patterns of cattle trade movements. PLoS ONE, [15] K. Berman. Vulnerability of scheduled networks and a gener-
6:e19869, 2011. alization of menger’s theorem. Networks, 28:125–134, 1996.
[5] P. Bajardi, A. Barrat, F. Natale, L. Savini, and V. Colizza. [16] B. Blonder and A. Dornhaus. Time-ordered networks reveal
Dynamical patterns of cattle trade movements. PLoS One, limitations to information flow in ant colonies. PLoS ONE,
6:e19869, 2011. 6:e20298, 2011.
[6] S. Bansal, J. Read, B. Pourbohloul, and L. A. Meyers. The [17] M. Boguñá, R. Pastor-Satorras, and A. Vespignani. Epidemic
dynamic nature of contact networks in infectious disease epi- spreading in complex networks with degree correlations. In
demiology. Journal of Biological Dynamics, 4:478–489, 2010. R. Pastor-Satorras, M. Rubi, and A. Diaz-Guilera, editors,
[7] A.-L. Barabási. The origin of bursts and heavy tails in humans Statistical mechanics of complex networks, pages 127–147.
dynamics. Nature, 435:207–212, 2005. Springer, Berlin, 2003.
[8] A. Barrat, M. Barthélemy, R. Pastor-Satorras, and A. Vespig- [18] D. Braha and Y. Bar-Yam. Time-dependent complex net-
nani. The architecture of weighted complex networks. Proc. works: dynamic centrality, dynamic motifs, and cycles of so-
Natl. Acad. Sci. USA, 101:3747, 2004. cial interaction. In T. Gross and H. Sayama, editors, Adap-
[9] A. Barrat, M. Barthélemy, and A. Vespignani. Dynamical tive networks: Theory, models and applications, pages 39–50.
processes on complex networks. Cambridge University Press, Springer, Dordrecht, 2008.
Cambridge UK, 2008. [19] B. Bui Xuan, A. Ferreira, and A. Jarry. Computing shortest,
[10] M. Barthélemy. Spatial networks. Physics Reports, 499:1– fastest, and foremost journeys in dynamic network. Interna-
101, 2011. tional Journal of Foundations of Computer Science, 14:267–
[11] M. Barthélemy, A. Barrat, R. Pastor-Satorras, and A. Vespig- 285, 2002.
nani. Velocity and hierarchical spread of epidemic outbreaks [20] E. Bullmore and O. Sporns. Complex brain networks: graph
in scale-free networks. Physical Review Letters, 92:178701, theoretical analysis of structural and functional systems. Na-
2004. ture Reviews Neuroscience, 10:186, 2009.
[12] D. S. Bassett, N. F. Wymbs, M. A. Porter, P. J. Mucha, J. M. [21] J. Candia, M. C. González, P. Wang, T. Schoenharl, G. Madey,
 25

and A.-L. Barabási. Uncovering individual and collective hu- [39] J.-P. Eckmann, E. Moses, and D. Sergi. Entropy of dialogues
man dynamics from mobile phone records. Journal of Physics creates coherent structures in e-mail traffic. Proc. Natl. Acad.
A, 41:224015, 2008. Sci. USA, 101:14333–14337, 2004.
[22] K. M. Carley. Dynamic network analysis. In R. Breiger, K. M. [40] A. Farrel. The Internet and its protocols: A comparative ap-
Carley, and P. Pattison, editors, Dynamic Social Network Mod- proach. Elsevier, Amsterdam, 2004.
eling and Analysis: Workshop Summary and Papers, pages [41] A. Ferreira. On models and algorithms for dynamic communi-
133–145. Committee on Human Factors, National Research cation networks: The case for evolving graphs. In Proceedings
Council, Washington DC, 2003. of 4e rencontres francophones sur les Aspects Algorithmiques
[23] A. Casteigts, P. Flocchini, W. Quattrociocchi, and N. Santoro. des Télécommunications (ALGOTEL‘2002), pages 155–161,
Time-varying graphs and dynamic networks. In Proceedings Mèze, 2002. INRIA Press.
of the 10th International Conference on Adhoc Networks and [42] S. Fortunato. Community detection in graphs. Physics Re-
Wireless (ADHOC-NOW), pages 346–359, 2011. ports, 486:75–174, 2010.
[24] C. Cattuto, W. van den Broeck, A. Barrat, V. Colizza, J.-F. [43] A. Gautreau, A. Barrat, and M. Barthélemy. Microdynamics
Pinton, and A. Vespignani. Dynamics of person-to-person in- in stationary complex networks. Proc. Natl. Acad. Sci. USA,
teractions from distributed RFID sensor networks. PLoS ONE, 106:8847–8852, 2009.
5:e11596, 2010. [44] S. Ghosh. Distributed Systems: An Algorithmic Approach.
[25] A. Chaintreau, A. Mtibaa, L. Massoulié, and C. Diot. Diame- Chapman & Hall / CRC, Boca Raton FL, 2007.
ter of opportunistic mobile networks. In Proceedings of ACM [45] K.-I. Goh and A.-L. Barabási. Burstiness and memory in com-
Sigcomm CoNext, 2007. plex systems. EPL, 81:48002, 2008.
[26] G. Chechik, E. Oh, O. Rando, J. Weissman, A. Regev, and [46] J. F. Gracia, V. M. Egu’iluz, and M. San Miguel. Update rules
D. Koller. Activity motifs reveal principles of timing in tran- and interevent time distributions: Slow ordering versus no or-
scriptional control of the yeast metabolic network. Nature dering in the voter model. Physical Review E, 84:015103,
Biotechnology, 26:1251–1259, 2008. 2011.
[27] E. Cheng, J. W. Grossman, and M. J. Lipman. Time-stamped [47] P. Grindrod, M. C. Parsons, D. J. Higham, and E. Estrada.
graphs and their associated influence digraphs. Discrete Ap- Communicability across evolving networks. Phys. Rev. E,
plied Mathematics, 128:317–335, 2003. 81:046120, 2011.
[28] A. Clauset and N. Eagle. Persistence and periodicity in a dy- [48] T. Gross and B. Blasius. Adaptive coevolutionary networks:
namic proximity network. In DIMACS Workshop on Computa- A review. J. Roy. Soc. Interface, 5:259–271, 2008.
tional Methods for Dynamic Interaction Networks, Piscataway [49] V. Gunturi, S. Shekhar, and A. Bhattacharya. Minimum span-
NJ, 2007. DIMACS. ning tree on spatio-temporal networks. In Proceedings of the
[29] R. Cohen, S. Havlin, and D. Ben-Avraham. Efficient im- 21th conference on database and expert systems application,
munization strategies for computer networks and populations. part 2., pages 149–158, Heidelberg, 2010. Springer.
Phys. Rev. Lett., 91:247901, 2003. [50] F. Guo, S. Hanneke, W. Fu, and E. P. Xing. Recovering tem-
[30] K. L. Cooke and E. Halsey. The shortest route through a net- porally rewiring networks: A model-based approach. In Inter-
work with time-dependent internodal transit times. Journal of national Conference of Machine Learning, 2007.
Mathematical Analysis and Applications, page 493, 1966. [51] S. Hachul and M. Jünger. An experimental comparison of fast
[31] D. P. Croft, J. Krause, and R. James. Social network in the algorithms for drawing general large graphs. Lecture Notes in
guppy (Poecilia reticulata). Proc. R. Soc. B., 271:S516–S519, Computer Science, 3843:235–250, 2006.
2004. [52] J.-D. J. Han, N. Bertin, T. Hao, D. S. Goldberg, G. F. Berriz,
[32] L. da Fontoura Costa, F. A. Rodriguez, G. Travieso, and P. Vil- L. V. Zhang, D. Dupuy, A. J. M. Walhout, M. E. Cusick,
las Boas. Characterization of complex networks: A survey of F. P. Roth, and . M. Vidali. Evidence for dynamically orga-
measurements. Advances in Physics, 56:167–242, 2007. nized modularity in the yeast proteinprotein interaction net-
[33] P. Dagum, A. Galper, and E. Horvitz. Dynamic network mod- work. Nature, 430:88–93, 2004.
els for forecasting. In Proceedings of the eighth conference on [53] S. Hanneke and E. P. Xing. Discrete temporal models of so-
Uncertainty in Artificial Intelligence, pages 41–48, 1992. cial networks. workshop on statistical network analysis. In
[34] P. C. de Ruiter, V. Wolters, and J. C. Moore, editors. Dynamic Proceedings of the 23rd International Conference on Machine
Food Webs: Multispecies Assemblages, Ecosystem Develop- Learning (ICML-SNA), 2006.
ment and Environmental Change. Academic Press, London, [54] F. Harary and G. Gupta. Dynamic graph models. Mathemati-
2005. cal and Computer Modelling, 25:79–88, 1997.
[35] F. de Vico Fallani, V. Latora, L. Astolfi, F. Cincotti, D. Mat- [55] T. E. Harris. The Theory of Branching Processes. Springer,
tia, M. G. Marciani, S. Salinari, A. Colosimo, and F. Babiloni. Berlin, 2002.
Persistent patterns of interconnection in time-varying cortical [56] H. W. Hethcote. The mathematics of infectious diseases.
networks estimated from high-resolution eeg recordings in hu- SIAM Review, 42:599, 2000.
mans during a simple motor act. J. Phys. A, 41:224014, 2008. [57] S. A. Hill and D. Braha. Dynamic model of time-dependent
[36] S. I. Dimitriadis, N. A. Laskaris, V. Tsirka, M. Vourkas, complex networks. Phys. Rev. E, 82:046105, 2010.
S. Micheloyannis, and S. Fotopoulos. Tracking brain dynam- [58] P. Holme. Network dynamics of ongoing social relationships.
ics via time-dependent network analysis. Journal of Neuro- Europhys. Lett., 64:427–433, 2003.
science Methods, 193:145, 2010. [59] P. Holme. Network reachability of real-world contact se-
[37] N. Eagle and A. Pentland. Reality mining: sensing complex quences. Phys. Rev. E, 71:046119, 2005.
social systems. Personal and Ubiquitous Computing, 10:255– [60] P. Holme, C. E. Edling, and F. Liljeros. Structure and time-
268, 2006. evolution of an Internet dating community. Social Networks,
[38] D. Easley and J. Kleinberg. Networks, crowds, and markets: 26:155–174, 2004.
reasoning about a highly connected world. Cambridge Uni- [61] J. L. Iribarren and E. Moro. Impact of human activity patterns
versity Press, Cambridge UK, 2010. on the dynamics of information diffusion. Phys. Rev. Lett.,
 26

103:038702, 2009. [83] F. Kuhn and R. Oshman. Dynamic networks: Models and
[62] J. L. Iribarren and E. Moro. Branching dynamics of viral in- algorithms. ACM SIGACT News, 42:82–96, 2011.
formation spreading. e-print arXiv:1110.1884, 2011. [84] R. Kumar, J. Novak, P. Raghavan, and A. Tomkins. On the
[63] L. Isella, M. Romano, A. Barrat, C. Cattuto, V. Colizza, bursty evolution of blogspace. In Proceedings of the 12th in-
W. Van den Broeck, F. Gesualdo, E. Pandolfi, L. Rav, ternational conference on World Wide Web, 2003.
C. Rizzo, and A. E. Tozzi. Close encounters in a pediatric [85] J. M. Kumpula, J. P. Onnela, J. Saramäki, K. Kaski, and
ward: Measuring face-to-face proximity and mixing patterns J. Kertész. Emergence of Communities in Weighted Networks.
with wearable sensors. PLoS ONE, 6:e17144, 2011. Physical Review Letters, 99:228701+, 2007.
[64] L. Isella, J. Stehlé, A. Barrat, C. Cattuto, J.-F. Pinton, and [86] Y. Kuwata, L. Blackmore, M. Wolf, N. Fathpour, C. Newman,
W. Van den Broeck. Whats in a crowd? analysis of face- and A. Elfes. Decomposition algorithm for global reachability
to-face behavioral networks. Journal of Theoretical Biology, analysis on a time-varying graph with an application to plan-
271:166–180, 2011. etary exploration. In International Conference on Intelligent
[65] M. O. Jackson. Social and economic networks. Princeton Robots and Systems, 2009.
University Press, Princeton NJ, 2008. [87] H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a
[66] A. Java, X. Song, T. Finin, and B. Tseng. Why we twitter: social network or a news media? In Proceedings of the 19th
Understanding microblogging usage and communities. In Pro- International World Wide Web Conference, 2010.
ceedings of the 9th WebKDD and 1st SNA-KDD workshop on [88] M. Lahiri and T. Y. Berger-Wolf. Structure prediction in tem-
web mining and social network analysis, 2007. poral networks using frequent subgraphs. In IEEE Symposium
[67] H.-H. Jo, M. Karsai, J. Kertész, and K. Kaski. Circadian pat- on Computational Intelligence and Data Mining, pages 35–
tern and burstiness in human communication activity. e-print 42, 2007.
arXiv:1101.0377. [89] M. Lahiri and T. Y. Berger-Wolf. Mining periodic behavior in
[68] H.-H. Jo, R. K. Pan, and K. Kaski. Emergence of bursts dynamic social networks. In Eighth IEEE International Con-
and communities in evolving weighted networks. PLoS ONE, ference on Data Mining, 2008.
6:e22687, 2011. [90] L. Lamport. Time, clocks, and the ordering of events in a
[69] A. Johansen. Probing human response times. Physica A, distributed system. Comm. ACM, 21:558–565, 1978.
330:286–291, 2004. [91] S. Lèbre. Inferring dynamic bayesian network with low or-
[70] C. Kamp. Untangling the interplay between epidemic spread der independencies. Statistical Applications in Genetics and
and transmission network dynamics. PLoS Comp. Biol., Molecular Biology, 8:9, 2009.
6:e1000984, 2010. [92] S. Lèbre, J. Becq, F. Devaux, M. P. H. Stumpf, and G. Le-
[71] M. Karsai, M. Kivel, R. K. Pan, K. Kaski, J. Kertész, A. L. landais. Statistical inference of the time-varying structure
Barabási, and J. Saramäki. Small but slow world: How net- of gene-regulation networks. BMC Systems Biology, 4:130,
work topology and burstiness slow down spreading. Phys. Rev. 2010.
E, 83:025102, 2011. [93] S. Lee, L. E. C. Rocha, F. Liljeros, and P. Holme. Exploit-
[72] J.-P. Kauppi, I. Jääskeläinen, M. Sams, and J. Tohka. Inter- ing temporal network structures of human interaction to effec-
subject correlation of brain hemodynamic responses during tively immunize populations. e-print arXiv:1011.3928.
watching a movie: localization in space and frequency. Jour- [94] W. E. Leland and D. V. Wilson. High time-resolution measure-
nal of Neuroinformatics, 4:5, 2009. ment and analysis of LAN traffic: Implications for lan inter-
[73] D. Kempe, J. Kleinberg, and A. Kumar. Connectivity and in- connection. In Proceedings of InfoCom91, pages 1360–1366,
ference problems for temporal networks. Journal of Computer 1991.
and System Sciences, 64:820, 2002. [95] J. Leskovec and E. Horvitz. Planetary-scale views on a large
[74] E. Kenah and J. M. Robins. Second look at the spread of epi- instant-messaging network. In Proceedings of the 17th Inter-
demics on networks. Physical Review E, 76:036113, 2007. national World Wide Web Conference, pages 915–924, 2008.
[75] M. Kimmel and D. E. Axelrod. Branching Process in Biology. [96] D. Liben-Nowell and J. Kleinberg. Tracing information flow
Springer, New York, 2002. on a global scale using Internet chain-letter data. Proc. Natl.
[76] J. Kleinberg. Bursty and hierarchical structure in streams. Acad. Sci. USA, 105:4633–4638, 2008.
Data Mining and Knowledge Discovery, 7:373–397, 2003. [97] F. Liljeros, C. E. Edling, and L. A. N. Amaral. Sexual net-
[77] M. Kolar, L. Song, A. Ahmed, and E. P. Xing. Estimating works: Implications for the transmission of sexually transmit-
time-varying networks. Annals of Applied Statistics, 4:94– ted infections. Microbes and Infections, 5:189196, 2003.
123, 2010. [98] F. Liljeros, C. E. Edling, L. A. N. Amaral, H. E. Stanley,
[78] K. Komurov and M. White. Revealing static and dynamic and Y. Åberg. The web of human sexual contacts. Nature,
modular architecture of the eukaryotic protein interaction net- 411:907–908, 2001.
work. Molecular Systems Biology, 3:110, 2007. [99] F. Liljeros, J. Giesecke, and P. Holme. The contact network of
[79] G. Kossinets, J. Kleinberg, and D. J. Watts. The structure of inpatients in a regional health care system: a longitudinal case
information pathways in a social communication network. In study. Mathematical Population Studies, 14:269–284, 2007.
Proc. 14th ACM SIGKKD Intl. Conf. on Knowledge Discovery [100] Y.-R. Lin, Y. Chi, S. Zhu, H. Sundaram, and B. L. Tseng.
and Data Mining, pages 435–443, 2008. Facetnet: a framework for analyzing communities and their
[80] V. Kostakos. Temporal graphs. Physica A, 388:1007–1023, evolutions in dynamic networks. In Proceedings of the 17th
2009. international conference on World Wide Web, pages 685–694,
[81] L. Kovanen, M. Karsai, K. Kaski, J. Kertész, and J. Saramäki. 2008.
Temporal motifs in time-dependent networks. e-print [101] D. Lusseau, K. Schneider, O. J. Boisseau, P. Haase, E. Slooten,
arXiv:1107.5646. and S. M. Dawson. The bottlenose dolphin community of
[82] M. Kretzschmar and M. Morris. Measures of concurrency in Doubtful Sound features a large proportion of long-lasting as-
networks and the spread of infectious disease. Math. Biosci., sociations. Behavioral Ecology and Sociobiology, 54:396–
133:165–195, 1996. 405, 2003.
 27

[102] R. D. Malmgren, D. B. Stouffer, A. S. L. O. Campanharo, and [125] R. Pastor-Satorras and A. Vespignani. Evolution and Structure
L. A. N. Amaral. On universality in human correspondence of the Internet: A Statistical Physics Approach. Cambridge
activity. Science, 325:1696–1700, 2009. University Press, Cambridge UK, 2004.
[103] R. D. Malmgren, D. B. Stouffer, A. E. Motter, and L. A. N. [126] T. M. Przytycka and M. S. D. K. Slonim. Toward the dy-
Amaral. A poissonian explanation for heavy tails in e-mail namic interactome: It’s about time. Briefings in Bioinformat-
communication. Proc. Natl. Acad. Sci. USA, 105:18153– ics, 11:15–29, 2010.
18158, 2008. [127] A. Rao, A. O. Hero III, D. J. States, and J. D. Engel. In-
[104] F. Mattern. Virtual time and global states of distributed sys- ferring time-varying network topologies from gene expression
tems. In Workshop on Parallel and Distributed Algorithms., data. EURASIP Journal on Bioinformatics and Systems Biol-
1989. ogy, 2007:51947, 2007.
[105] M. Medo, G. Cimini, and S. Gualdi. Temporal effects in the [128] C. S. Riolo, J. S. Koopman, and J. S. Chick. Methods and mea-
growth of networks. e-print arXiv:1009.5560. sures for the description of epidemiological contact networks.
[106] B. Min, K.-I. Goh, and A. Vazquez. Spreading dy- Journal of Urban Health, 78:446–457, 2001.
namics following bursty human activity patterns. e-print [129] G. Robins, P. Pattison, Y. Kalish, and D. Lusher. An introduc-
arXiv:1006.2643. tion to exponential random graph models for social networks.
[107] G. Miritello, E. Moro, and R. Lara. The dynamical strength of Social Networks, 29:173–191, 2006.
social ties in information spreading. Phys. Rev. E, 83:045102, [130] L. E. C. Rocha, F. Liljeros, and P. Holme. Information dy-
2011. namics shape the sexual networks of internet-mediated prosti-
[108] J. Moody. The importance of relationship timing for diffusion. tution. Proc. Natl. Acad. Sci. USA, 107:5706–5711, 2010.
Social Forces, 81:25–56, 2002. [131] L. E. C. Rocha, F. Liljeros, and P. Holme. Simulated epidemics
[109] M. Morris and M. Kretzschmar. Concurrent partnerships and in an empirical spatiotemporal network of 50,185 sexual con-
transmission dynamics in networks. Social Networks, 17:299– tacts. PloS Comp. Biol., 7:e1001109, 2011.
318, 1995. [132] P. Ronhovde, S. Chakrabarty, D. Hu, M. Sahu, K. F. Kelton,
[110] P. J. Mucha, T. Richardson, K. Macon, M. A. Porter, and J.-P. N. A. Mauro, K. K. Sahu, and Z. Nussinov. Detecting hidden
Onnela. Community structure in time-dependent, multiscale, spatial and spatio-temporal structures in glasses and complex
and multiplex networks. Science, 328:876–878, 2010. physical systems by multiresolution network clustering. 2011.
[111] M. E. J. Newman. Spread of epidemic disease on networks. eprint arXiv:1102.1519.
Phys. Rev. E, 66:16128, 2002. [133] M. Rosvall and C. T. Bergstrom. Mapping change in large
[112] M. E. J. Newman. Networks: An introduction. Oxford Uni- networks. PLoS ONE, 5:e8694, 2010.
versity Press, Oxford UK, 2010. [134] N. Santoro, W. Quattrociocchi, P. Flocchini, A. Casteigts, and
[113] V. Nicosia, J. Tang, M. Musolesi, G. Russo, C. Mascolo, F. Amblard. Time-varying graphs and social network anal-
and V. Latora. Components in time-varying graphs. e-print ysis: temporal indicators and metric. In Proceedings of the
arXiv:1106.2134. 3rd AISB Social Networks and Multiagent Systems Symposium
[114] M. K. Nordvik and F. Liljeros. Number of sexual encounters (SNAMAS), pages 32–38, 2011.
involving intercourse and the transmission of sexually trans- [135] T. A. B. Snijders, J. Koskinen, and M. Schweinberger. Maxi-
mitted infections. Sexually Transmitted Diseases, 33:342–349, mum likelihood estimation for social network dynamics. The
2006. Annals of Applied Statistics, 4:567–588, 2010.
[115] J. G. Oliveira and A.-L. Barabási. Human dynamics: Dar- [136] T. A. B. Snijders, G. G. van de Bunt, and C. E. G. Steglich.
win and Einstein correspondence patterns. Nature, 437:1251, Introduction to stochastic actor-based models for network dy-
2005. namics. Social Networks, 32:44–60, 2010.
[116] J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, D. Lazer, [137] R. V. Solé and J. Bascompte. Self-Organization in Complex
K. Kaski, J. Kertész, and A.-L. Barabási. Structure and tie Ecosystems. Princeton University Press, Princeton NJ, 2006.
strengths in mobile communication networks. Proc. Natl. [138] O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilgetag. Or-
Acad. Sci. USA, 104:7332, 2007. ganization, development and function of complex brain net-
[117] C. Pahl-Wostl. The dynamic nature of ecosystems: chaos and works. Trends in Cognitive Sciences, 8:418–425, 2004.
order entwined. Wiley, Chichester UK, 1995. [139] J. Stehlé, A. Barrat, and G. Bianconi. Dynamical and bursty
[118] G. Palla, A.-L. Barabási, and T. Vicsek. Quantifying social interactions in social networks. Phys. Rev. E, 81:035101,
group evolution. Nature, 446:664–667, 2007. 2010.
[119] B. Ø. Palsson. Systems Biology: Properties of Reconstructed [140] J. Stehlé, N. Voirin, A. Barrat, C. Cattuto, V. Colizza, L. Isella,
Networks. Cambridge University Press, Cambridge UK, 2006. C. Regis, J.-F. Pinton, N. Khanafer, W. Van den Broeck, and
[120] R. K. Pan and J. Saramäki. Path lengths, correlations, and cen- P. Vanhems. Simulation of an SEIR infectious disease model
trality in temporal networks. Phys. Rev. E, 84:016105, 2011. on the dynamic contact network of conference attendees. BMC
[121] A. Panisson, A. Barrat, C. Cattuto, W. Van den Broeck, Medicine, 9(87), 2011.
G. Ruffo, and R. Schifanella. On the dynamics of human prox- [141] J. Stehlé, N. Voirin, A. Barrat, C. Cattuto, V. Colizza, L. Isella,
imity for data diffusion in ad-hoc networks. to appear in Ad C. Regis, J.-F. Pinton, N. Khanafer, W. Van den Broeck, and
Hoc Networks, 2011. P. Vanhems. Simulation of an seir infectious disease model on
[122] Y. Park, C. Moore, and J. S. Bader. Dynamic networks from the dynamic contact network of conference attendees. BMC
hierarchical Bayesian graph clustering. PLoS ONE, 5:e8118, Medicine, 9:87, 2011.
2010. [142] J. Stehlé, N. Voirin, A. Barrat, C. Cattuto, L. Isella, J.-F. Pin-
[123] M. Pascual and J. Dunne. Ecological Networks: Linking ton, M. Quaggiotto, W. Van den Broeck, C. Rgis, B. Lina, and
Structure to Dynamics in Food Webs. Oxford University Press, P. Vanhems. High-resolution measurements of face-to-face
Oxford UK, 2006. contact patterns in a primary school. PLoS ONE, 6:e23176,
[124] R. Pastor-Satorras and A. Vespignani. Epidemic spreading in 2011.
scale-free networks. Phys. Rev. Lett., 86:3200–3203, 2001. [143] S. R. Sundaresan, I. R. Fischhoff, J. Dushoff, and D. I. Ruben-
 28

stein. Network metrics reveal differences in social organiza- epidemics in dynamic contact networks. Proc. Roy. Soc. B,
tion between two fission-fusion species, Grevy’s zebra and on- 274:2925–2934, 2007.
ager. Oecologia, 151:140–149, 2006. [160] S. Wasserman and K. Faust. Social Network Analysis: Meth-
[144] B. Szendroi and G. Csanyi. Polynomial epidemics and cluster- ods and Applications. Cambridge University Press, Cam-
ing in contact networks. Proc. Roy. Soc. B, 271:S364–S366, bridge UK, 1994.
2004. [161] D. J. Watts and S. H. Strogatz. Collective dynamics of ‘small-
[145] T. Takaguchi, M. Nakamura, N. Sato, K. Yano, and N. Ma- world’ networks. Nature, 393:409–410, 1998.
suda. Predictability of conversation patterns. Phys. Rev. X, [162] Y. Wu, C. Zhou, J. Xiao, J. Kurths, and H. J. Schellnhuber.
1:011008, 2011. Evidence for a bimodal distribution in human communication.
[146] J. Tang, C. Mascolo, M. Musolesi, and V. Latora. Exploring Proc. Natl. Acad. Sci. USA, 107:18803–18808, 2010.
temporal complex network metrics in mobile malware con- [163] Z. Yang, A.-X. Cui, and T. Zhou. Impact of heterogeneous hu-
tainment. In 12th IEEE International Symposium on a World man activities on epidemic spreading. Physica A, 390:4543–
of Wireless, Mobile and Multimedia Networks (WOWMOM 4548, 2011.
‘11), 2011. [164] T. Yasseri, R. Sumi, and J. Kertész. Circadian patterns of
[147] J. Tang, M. Musolesi, C. Mascolo, and V. Latora. Temporal Wikipedia editorial activity: A demographic analysis. eprint
distance metrics for social network analysis. In Proceedings arXiv:1109.1746, 2011.
of the 2nd ACM SIGCOMM Workshop on Online Social Net- [165] R. Yoshida, S. Imoto, and T. Higuchi. Estimating time-
works, page 3, 2009. dependent gene networks from time series microarray data by
[148] J. Tang, M. Musolesi, C. Mascolo, V. Latora, and V. Nicosia. dynamic linear models with markov switching. In CSB ’05:
Analysing information flows and key mediators through tem- Proceedings of the 2005 IEEE Computational Systems Bioin-
poral centrality metrics. In Proceeding of the 3rd ACM Eu- formatics Conference, 289-298, 2005. IEEE Computer Soci-
roSys Workshop on Social Networks Systems (SNS’10), page 3, ety.
2010. [166] T. Yoshida, L. E. Jones, S. P. Ellner, G. F. Fussmann, and N. G.
[149] J. Tang, S. Scellato, M. Musolesi, C. Mascolo, and V. Latora. Hairston Jr. Rapid evolution drives ecological dynamics in a
Small-world behavior in time-varying graphs. Phys. Rev. E, predator-prey system. Nature, 424:303–306, 2003.
81:055101, 2010. [167] K. Zhao, J. Stehlé, G. Bianconi, and A. Barrat. Social net-
[150] C. Tantipathananandh, T. Y. Berger-Wolf, and D. Kempe. A work dynamics of face-to-face interactions. Phys. Rev. E,
framework for community identification in dynamical social 83:056109, 2011.
networks. In Proceedings of the 13th ACM SIGKDD Interna- [168] Q. Zhao, Y. Tian, Q. He, N. Oliver, R. Jin, and W.-C. Lee.
tional Conference on Knowledge Discovery and Data Mining, Communication motifs: A tool to characterize social com-
pages 717–726, 2007. munications. In Proceedings of the 19th ACM international
[151] I. W. Taylor, R. Linding, D. Warde-Farley, Y. Liu, C. Pesquita, conference on Information and knowledge management, page
D. Faria, S. Bullamd, T. Pawson, Q. Morris, and J. L. Wrana. 1645, 2010.
Dynamic modularity in protein interaction networks predicts [169] Z.-D. Zhao, H. Xia, M.-S. Shang, and T. Zhou. Empirical
breast cancer outcomes. Nature Biotech., 27:199–204, 2009. analysis on the human dynamics of a large-scale short message
[152] R. Timo, K. Blackmore, and L. Hanlen. On entropy mea- communication system. Chinese Physics Letters, 28:068901,
sures for dynamic network topologies: Limits to MANET. In 2011.
Proceedings of the sixth Australian Communications Theory [170] T. Zhou, Z.-D. Zhao, Z. Yang, and C. Zhou. Rela-
Workshop, pages 95–101, 2005. tiveclockverifiesendogenousburstsofhumandynamics. e-print
[153] T. S. Turova. Dynamical random graphs with memory. Phys. arXiv:1106.5562.
Rev. E, 65:066102, 2002. [171] Kostakos defines a set of measures for generation graphs, in
[154] T. Ueno and N. Masuda. Controlling nosocomial infection increasing specificity (some corresponding to latency, some
based on structure of hospital social networks. J. Theor. Biol., corresponding to the shortest latency over a time interval)
254:655–666, 2008. which he all calls temporal proximity.
[155] R. E. Ulanowicz. Quantitative methods for ecological network [172] It should be noted that the word motif is often used in
analysis. Comp. Biol. Chem., 28:321–339, 2004. other meanings than an overrepresented class of subgraphs—
[156] M. Valencia, J. Martinerie, S. Dupont, and M. Chavez. Dy- equivalence classes such as feedforward triangles may be
namic small-world behavior in functional brain networks un- called motifs whether they are overrepresented or not. In ad-
veiled by an event-related networks approach. Phys. Rev. E, dition, the word motif is at times used to denote the subgraphs
77:050905, 2008. that constitute a motif
[157] A. Vazquez, B. Rácz, A. Lukács, and A.-L. Barabási. Impact [173] Note that in some contexts, compartmental models refer to
of non-poissonian activity patterns on spreading processes. models where for each state, one has one variable counting
Phys. Rev. Lett., 98:158702, 2007. the size of the population in that state, instead of models that
[158] M. C. Vernon and M. J. Keeling. Representing the UK’s cat- explicitly deal with individuals like in the models discussed
tle herd as static and dynamic networks. Proc. R. Soc. B, here.
276:469–476, 2009.
[159] E. Volz and L. A. Meyers. Susceptible-infected-recovered